"POLITEHNICA" UNIVERSITY OF TIMISOARA ROMANIA UNIVERSITÉ DE FRANCHE-COMTÉ BESANÇON FRANCE AUTOMATICS and COMPUTER SCIENCE

Department :

Computer Sciences

Medical Image Processing

HŸY Compliant for PD Diagnosis and Prognosis using EPI and FA Images P H D R E P O R T No. 2

PhD Student

Roxana Oana Teodorescu

Thesis Advisor: Vladimir Ioan Cretu and Daniel Racoceanu

"Politehnica" University of Timisoara and Image Ÿ Pervasive Access Lab Singapore

February 4, 2010

Design and Use of DTI Medical Images for PD Diagnosis Abstract: My work is focused in nding solutions for medical image processing

and analysis in MRI images with applications for Parkinson's Disease. At the processing level I develop algorithms that are independent on the patient variability for detecting the volumes of interest, fusing dierent image information type. A rigid registration with automatic detection of the geometrical parameters makes possible the fusion step, by eliminating the volume variability from the analysis step. The analysis is possible by tracking the strationigral bers even in the gray matter. We study Parkinson's disease (PD) using an automatic specialized diusion-based atlas. A total of 47 subjects, among who 22 patients diagnosed clinically with PD and 25 control cases, underwent DTI imaging. The EPIs have lower resolution but provide essential anisotropy information for the ber tracking process. The two volumes of interest (VOI) represented by the Substantia Nigra and the Putamen are detected on the EPI and FA respectively. We use the VOIs for the geometry-based registration. We fuse the anatomical detail detected on FA image for the putamen volume with the EPI. After 3D bers growing on the two volumes, we compute the ber density (FD) and the ber volume (FV). Furthermore, we compare patients based on the extracted bers and evaluate them according to Hohen&Yahr (H&Y) scale. This report introduces the method used for automatic volume detection and evaluates the ber growing method on these volumes. Our approach is important from the clinical standpoint, providing a new tool for the neurologists to evaluate and predict PD evolution. From the technical point of view, the fusion approach deals with the tensor based information (EPI) and the extraction of the anatomical detail (FA and EPI). Keywords: Automatic ROI/VOI detection, Medical Image Analysis, Medical Image Processing, PD Detection, Prediction

Acknowledgments The work presented in this Report has been the result of a collaboration with Dr. Ling-Ling Chan and her team from Singapore General Hospital (SGH). For all the brain anatomy lessons and the image, as well as for her work for gathering the database and make it available for us. At the beginning we have collaborated with Dr. Karl Olof Lövblad from the Universities Hospitals of Geneva and I want to thank him for his time and patience. The stage made in Singapore from February until October 2009 has been a real help in my research and I want to thank the IPAL1 team, NUS2 and UPT3 for this. This stage has been supported by CNRS 4 and the CNCSIS 5 scholarship TD6 46/2008 from the Romanian government.

1

Image and Pervasive Access Lab - http://ipal.i2r.a-star.edu.sg/ National University of Singapore - http://www.nus.edu.sg/ 3 "Politehnica" University of Timisoara -http://www.cs.upt.ro 4 French National Research Center

2

5 6

http://www.cncsis.ro/

TD - Young PhD students Scholarship

Contents

1 Introduction 1.1

1.2 1.3 1.4

Motivation. Purpose. Idea . . . . . . . . . . . . . . . . . . . 1.1.1 Current diagnosis for Parkinson's Disease . . . . . . 1.1.2 Challenges. Our Solutions. . . . . . . . . . . . . . . . 1.1.3 Proposed Approach: introducing PDAtl@s . . . . . . Idea behind the system - general presentation of the system Report structure . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Database and its importance 2.1 2.2

2.3 2.4 2.5 2.6

Diusion Tensor Images (DTI) . . . . . . . . . . . . . Echo-planar images (EPI) . . . . . . . . . . . . . . . . 2.2.1 Tensors and their purpose . . . . . . . . . . . . 2.2.2 Managing the information coming from the EPI Fractional Anisotropy (FA) images . . . . . . . . . . . Other types of DTI images and their importance . . . Preparing the image for processing . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

3 Our Approach on Image Processing and Analysis 3.1

3.2 3.3 3.4 3.5 3.6 3.7 3.8

Overview on the system . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.1 Comparison with other type of images and correlation with current diagnoses system . . . . . . . . . . . . . . . . . . . . . 3.1.2 Proposed approach . . . . . . . . . . . . . . . . . . . . . . . . 3.1.3 PDFibAtl@s : our system for diagnoses . . . . . . . . . . . . Removing the artifacts & the skull . . . . . . . . . . . . . . . . . . . Preliminary testing. Motivation for further study. . . . . . . . . . . . Detecting the Volumes of Interest (VOIs) . . . . . . . . . . . . . . . 3.4.1 Automatic detection of the volume of interest (VOI) . . . . . Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Fusion factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Growing bers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4 Validation, Results and Interpretation 4.1

4.2

Medical relevance . . . . . . . . . . . . . . . . . 4.1.1 Green Channel analysis on the midbrain 4.1.2 Fiber study . . . . . . . . . . . . . . . . Technical specic elements . . . . . . . . . . . . 4.2.1 Medical Image processing . . . . . . . .

. . . area . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

1 2 3 3 5 6 6 6

9

9 11 11 11 11 12 13 14

15

15

16 17 18 20 20 22 23 26 28 28 29

31

32 32 33 36 36

vi

Contents 4.3

4.2.2 Speed for computation . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5 Conclusion. Further study. Applicability 5.1 5.2 5.3 5.4 5.5

Diagnosis evaluation . . . . . . . . . . . . . . Prognosis potential . . . . . . . . . . . . . . . PDFibAtl@s importance. Future applicability. Contribution . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

37 37

39

39 40 40 41 42

List of Abbreviations

43

A Appendix Dissemination

45

Bibliography

47

A.1 Conferences & Workshops 2009 -2010 . . . . . . . . . . . . . . . . . . A.2 Research stages 2009-2010 . . . . . . . . . . . . . . . . . . . . . . . .

45 45

List of Figures

2.1 2.2 2.3

Database input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . EPI and FA example images . . . . . . . . . . . . . . . . . . . . . . . T1 and T2 example images . . . . . . . . . . . . . . . . . . . . . . .

10 12 13

3.1 3.2 3.3 3.4 3.5

. . . . . . . . . . . . . . . . bers . . . . brain . . . . . . . .

19 20 21 22

3.7

Layout of PDFibAtl@s . . . . . . . . . . . . . . . . . . . . . . Image of the brain contour . . . . . . . . . . . . . . . . . . . . 3D view of EPI stack with B0 images . . . . . . . . . . . . . . Green channel analysis . . . . . . . . . . . . . . . . . . . . . . EPI with detected VOIs in 3.5(a), 3.5(b) and 3.5(c) with 3D on an example . . . . . . . . . . . . . . . . . . . . . . . . . . The main anatomical structures at the putamen level of the [Talos 2003] . . . . . . . . . . . . . . . . . . . . . . . . . . . . Geometrical view of the registration parameters . . . . . . . .

4.1 4.2

The Results Window [Teodorescu 2009d] [Teodorescu 2009a] . . . . . 3D View of the grown bers . . . . . . . . . . . . . . . . . . . . . . .

32 34

3.6

24 25 27

List of Tables

4.1 4.2 4.3 4.4 4.5

Test batches characteristics [Teodorescu 2009b] . . . . . . . . . . . . Study on Green channel on the left and right side [Teodorescu 2009b] Simple correlation between the FV and H&Y values[Teodorescu 2010] ANOVA testing [Teodorescu 2010] . . . . . . . . . . . . . . . . . . . Variation of density [Teodorescu 2010] . . . . . . . . . . . . . . . . .

33 33 34 35 35

Chapter 1 Introduction

Contents 1.1 Motivation. Purpose. Idea . . . . . . . . . . . . . . . . . . . . 1.1.1 1.1.2 1.1.3

Current diagnosis for Parkinson's Disease . . . . . . . . . . . Challenges. Our Solutions. . . . . . . . . . . . . . . . . . . . Proposed Approach: introducing PDAtl@s . . . . . . . . . . .

2 3 3 5

1.2 Idea behind the system - general presentation of the system 6 1.3 Report structure . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 As the present report contains a continuation of the thesis report Nr.1 [Teodorescu 2009c], I am making an overview of the main purpose of the thesis in this chapter presenting the challenges that I deal with and the ideas used for each problem. A short overview of the ideas from the rst report is included as well by mentioning the test performed and the importance of the results at that point. The research ideas presented there inuenced the ideas from the present inform. This chapter constitutes the link between the two reports and an overview of the proposed system. In [Teodorescu 2009c] I have made a study on the specic characteristics of the medical images, types of medical images currently used by the medical doctors, as well as technical details of the dierent protocol procedures that can be applied for PD. A special place is given to the presentation of the Digital Imaging and Communications in Medicine(DICOM) standard, as it is used for the images in our database. The Analyze standard for medical images is presented as well, because we use it in the processing stage. For the Magnetic Resonance Images (MRI) images, with detailed presentation of the Diusion Tensor Images(DTI) protocol, I included also a motivation for using this type of images in our application. Specic metrics for PD in the DTI images are introduced, as well as a global analysis of the database. The image quality for our database is evaluated in the same report by extracting the introduced metrics (FA/ADC) . Only after an evaluation of the images we can evaluate the problems and envision a competent approach for solving them by creating new algorithms at the image processing level. The existing systems that try to solve the same problems as ours are evaluated in the present report, together with our own approach for comparison purpose. PDFibAtl@s is the system that we have created to accomplish a diagnosis analysis based on 3D image processing. It retrieves the 3D volume image of the brain from

2

Chapter 1. Introduction

medical format, it treats the volume by examining the interesting areas, extracts the motor bers and computes their density and the relative volume. All constitutive algorithms of the proposed system are the result of our research work started in 2008 and presented as preliminary proves in the PhD Report No.1. This report contained the state of the art on PD and the testing performed of the medical images and their specic characteristics. The medical images were employed to determine the fractional anisotropy and the apparent diusion coecient level in order to see the relevance of these indicator relative to PD. These tests proved to be meaningful as we are using for our system the FA image now. These images are obtained from the fractional anisotropy of the voxels. The preceding report constituted the basics for the development of PDFibAtl@s by providing the study on the optimal imaging types needed for our purpose. Even if the bare results on the computed coecients presented in the previous report did not show a clear limitation between the patients and the control cases due to the inclusion of the skull and the vast volume of study, it provided the primary data for the system. We needed an image of the information contained in the medical images in our database and fundamental overview of the PD problems. A correlation between the images and the PD is obtained only by nding the signs of the disease in the images. This study represented the focus of the rst report and a solution for determining automatically and integrating the indicators of PD as an image-system represents the purpose of the present report. For this report we have delimited the volume of study and we detect the volume of interest as being the source of the dopamine, where the neural bers from the motor tract are growing. These are the main indicators, together with the cognitive tests, for the installation of the disease on a subject. This report is based on the global study of the brain performed on the previous report, but it concentrates on specic areas of the brain, developing methods to determine these areas automatically, to analyze the images providing the medical knowledge and to make a correlation between the disease and the image characteristics. This report presents the way that our system detects the relevant features for PD and based on these features we develop a diagnosis and a prognosis system.

1.1 Motivation. Purpose. Idea Parkinson's disease aects population that has, on average, 61 years, even if it begins around 40 years[PD 2009]. From this point of view, the continuous aging of the population, combined with the actual late detection (impossibility to reverse or stabilize the PD evolution) justies concerns for a prediction system. By the time the disease is detected, the patient has already lost 80-90% of the dopamine cells [news today 2009], those that represent one of the main neurotransmitters. The treatments are less eective after the disease develops. Thus a prognosis of this disease could diminish the eect of the PD or even reverse it. Analyzing the current modality of diagnoses, our idea aims at augmenting the trust degree on the diagnosis by adding the image in the process of detection of the disease and making possible

1.1. Motivation. Purpose. Idea

3

the detection at an early stage as well - the prognosis factor.

1.1.1

Current diagnosis for Parkinson's Disease

Currently the procedure for diagnoses of the PD patients relies in cognitive testing for the patients and according to their scores they are placed on a predened scale: Unied Parkinson Rating Scale or H&Y (Hohen & Yahr scale). This manner of diagnosis does not take into account the information provided by the images. Performing an analysis of the images and nding an association between the eect of the disease represented by image specic indicators we can integrate this perspective in the diagnose decision as well. The advantage in this case is given by compounding the cognitive aspect with the anatomical one. Our medical team provided us with the H&Y scale for all the patients as a ground truth for the relation between the parameters computed and the PD severity. According to the study performed by Dr. Chan, there is a match between the dopamine level in the Substantia Nigra and the Parkinson's disease evolution [Chan 2007]. The study takes into account the manually detected area where the Substantia Nigra(SN) is supposed to be anatomically placed. This segmented area is further studied to determine the correlation between the PD patients and the dopamine level in this area. A correlation has been found indeed, but to make the dierence for diagnoses purposes, this correlation is not enough and not reliable. David Vaillancourt, assistant professor at UIC has scanned the part of the brain called Substantia Nigra on Parkinson's patients using DTI images and has discovered that the number of dopaminergic neurons in certain areas of this region is 50% less [Vaillancourt 2009]. His study includes 28 subjects from which half have symptoms of early Parkinson's disease and another half do not have these symptoms. A study performed to show the relationship between cerebral morphology and the expression of dopamine receptors conducted on 45 healthy patients, reveals that on grey matter images, there is a direct correlation at the SN level. This study [Woodward 2009] uses T1 weighted structural MRI images. Using Voxel-based morphometry (VBM) the authors create grey matter volumes and density images and correlate these images with Biological Parametric toolbox. Voxel-wise normalization revealed also that the grey matter volume and SN are density positively correlated. On the brain ber tracts the dopamine should ow from the back of the brain toward the front and from up towards the basis of the brain. In [Lehericyr 2004] a manual detection of the regions of interest in the basal ganglia area has been operated and some dierences are observed.

1.1.2

Challenges. Our Solutions.

One of the main eects of Parkinson is represented by the lost of mobility perceived as a trembling eect on the patients. We study the motor tracts in order to determine if there is a direct link on the lost of dopamine and the degeneration of the neural bers of this tract. A statistical analysis of the number of bers and their density

4

Chapter 1. Introduction

is able to reveal if there is a relationship between the bers and the PD severity . In order to identify the right bundle of bers from the motor tract, we choose the starting area and the ending point of the bers such that we can eliminate all the bers that do not cross the two areas in the same time. As the dopamine is produced in the Substantia Nigra area, the starting point for these bers is represented by this anatomical region. Then we choose the Putamen area as the ending point for the selected bers. The medical area, by dening the specic problems, provided us with the technical challenges in the anatomical details and the images used. The fact that each brain structure is dierent from one patient to the other (e.g the placement of the putamen with regard to the center of mass of the brain, the size of it) represents a challenge from the specicity point of view for each medical case. A brain image has a dierent aspect, just like the portrait photo dierers from one person to another. This is the reason why determining a certain anatomical aspect from a brain image is similar to determining the nose position or the mouth in a portraits database. Detecting an anatomical volume automatically on a medical image can be compared with the same challenge on a 3D image of a person. While there are several algorithms that determine the position of the eyes and/or the mouth - face recognition specic algorithms-, we cannot say the same things for the brain images. Atlases for the brain have the same meaning as image patterns applied for portraits. On the portraits all the anatomical elements are known, but not the same things are applicable for the brain images - there are formations that can be placed in dierent positions with respect to the center of mass of the image or of the brain; dierent sizes with regard to the other elements inside the same brain. This is the reason why an atlas applied like a pattern or a mask on a brain will not necessarily give a correct result when detecting a specic anatomical element. A specialized algorithm for detecting an anatomical volume inside a brain, that does not rely on the specicity of an image is much needed at this point. This kind of algorithm must have robustness and be exact as well on determining the needed elements. Technically this represents a big challenge and relies on nding undeviating points in the anatomy of the brain, as well as relative positions for the searched areas. There are special limitations regarding the medical images resolution and specicity for such an algorithm. One of the main tasks is to nd the appropriate slice in which to look for the volume of interest. Each slice contains a dierent information and we relay on volumetric information when choosing the slice of interest for each of the segmentation algorithms. Another need at this point is to nd the region where we apply the segmentation and while for the midbrain is easier as we take the center of mass of the brain in the slice of interest that is contained in its volume, for the putamen is rather dicult. This anatomical region is not symmetrically placed on the left and right side of the middle axis that separates the hemispheres, neither at the same relative position with regard to the center of mass. This is one of the challenges together with the fact that the right side putamen can have a dierent shape and size from the left side and be placed higher or lower than the other one. Finding tough the midline that delimits the two hemispheres of the brain is another

1.1. Motivation. Purpose. Idea

5

bid as it must be determined. The two hemispheres are not symmetrical and the line is not necessarily perpendicular on the horizontal axis of the image. I need to determine this axis with no connection to the specicity of the patient, I need an independent way to determine it. The specic algorithms that detect the volumes of interest take into account the voxel intensity and are dened and presented in Chapter 3. Another major objective is represented by the registration aspect. As the T1/T2 images have high resolution, but the intensity of the pixels does not permit an accurate limitation for my detection algorithm, I apply it on the FA image. This provides an accurate result but needs to be registered with the EPI image for further use in the ber growth algorithm. Although the registration algorithm by itself does not represent a new approach, the detection of the parameters for applying this algorithm does. I make the detection for the algorithm by using the geometrical parameters, like the middle hemisphere axis, determined by an original approach as well. The nal aim in building PDFibAtl@s is represented by the limitation of the grown bers. The algorithm that we use does not make any dierence between the tensors, the anatomical placement of the bers and the tract that they belong to. I need for PD just the motor tract and I am using the detected volumes to select the bers that pass through both volumes of interest so that I am analyzing just the needed tract.

1.1.3

Proposed Approach: introducing PDAtl@s

Our program is able to automatically detect on the Echo-Planar Images (EPI) the slice where the midbrain area is located, where the Substantia Nigra resides. Substantia Nigra represents the area of the brain where the dopamine is produced. This area is not anatomically dened in the specialized atlases and that is why I identify the midbrain area, as it contains the Substantia Nigra. The Putamen area is now well dened on the EPI images, or in the T2 images, but the FA images have a clear boundary of the Putamen area. We use these images to locate the specied region of interest. Working with 3D images, I identify two volumes of interest represented by the midbrain (2 slices) and the Putamen (3 slices) by using several slices each time. I start the growing process from the midbrain towards the Putamen. The bers selected by the program are then statistically analyzed and using the T-Test we examine the correlation between the coecients determined on the bers and the evolution of the disease.

6

Chapter 1. Introduction

1.2 Idea behind the system - general presentation of the system Analyzing the images taken form PD patients graded on the H&Y scale we grow the bers from the motor tract and study the relation between the situation of the bers and the rated values. After this analysis as the identication of the volumes of interest is automatic, we are relating to an automatic dedicated atlas for PD regions. Making a variation function to show the PD evolution, based on the bers coecients, by extrapolating this function we can determine the values of the coecients for the early PD cases. This extrapolation factor gives a prognosis value that can be used and tested on the early PD cases. The approach shown in this report includes a new automatic detection algorithm for the midbrain area together with a geometric approach for detecting the putamen as a basis for a fully automated atlas of the brain. I modify an existing algorithm for ber growth by limiting the bers using the automatically detected volumes. The metrics introduced here for the bers are our own. All the research that leaded to the realization of PDFibAtl@s with a presentation of the algorithms is the aim of this report.

1.3 Report structure Representing the progress of the work from the whole brain analysis presented on the rst report, I move towards determining the neural bers from the motor tract and this design with all the steps are emphasized in the present report. The image analysis and the specicity of the image types used are presented in chapter 2. Than the algorithms used for the image processing and extracting the concepts are detailed in 3 and referred by the validation and results for each step of the system in 4. An overview of the entire work and the nal conclusions are presented in chapter 5.

1.4 Conclusion The purpose of our work is to detect and predict the evolution of the PD. We implement several algorithms for detection of the volume of interest and registration in order to integrate a fusion of information from the two types of modalities involved: EPI and FA. On the volumes of interest detected we grow the bers and using their density we analyze the eects of the disease. Our system performs the image processing part, the analysis part being performed using a statistical tool. Considering the medical needs we have several problems to solve:

• Automatic detection of volumes of interest • Fusion of image information • Medical image registration

1.4. Conclusion

7

• Fiber growing through the detected volumes of interest • Finding the appropriate coecients to evaluate the conditions of the bers These are the main objectives for us from the technical point of view. To reach these goals we need several algorithms for DICOM header management and medical image preprocessing.

Chapter 2 Database and its importance

Contents 2.1 Diusion Tensor Images (DTI) . . . . . . . . . . . . . . . . . 9 2.2 Echo-planar images (EPI) . . . . . . . . . . . . . . . . . . . . 11 2.3 2.4 2.5 2.6

2.2.1 2.2.2

Tensors and their purpose . . . . . . . . . . . . . . . . . . . . Managing the information coming from the EPI . . . . . . .

Fractional Anisotropy (FA) images . . . . . . . . . Other types of DTI images and their importance Preparing the image for processing . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

11 11

11 12 13 14

There are several aspects we must take into account when dening the algorithms for the implementation of PDFibAtl@s. As the images represent the input, their quality inuence the overall results and the approach used for the algorithms. Given the images provided by our medical team, we further consider several technical facts related to the specicity of the medical images. In order to have the best results on the detection and analysis performed on the images, the quality of these images has to compel with the needs of our algorithms: good resolution and anatomical detail. Another important aspect is represented by the patient identity factor that interferes with our analysis: each patient has a specic anatomy and the slice of interest for dierent patients is not the same one. As the volume of the brain is not the same for all the patients and the anatomical detail diers among the cases, the automatic detection of specic areas should be based on the geometrical display of the brain anatomy. In addition, a correlation between the patient's age and the sex must be taken into account since we should exclude the atrophy and other geriatric specic elements, as well as dierences in the volume of the brain due to sex dierentiation. This chapter presents the medical image characteristics specic for the DTI images emphasizing on the images that we are using and the specic characteristics of these images.

2.1 Diusion Tensor Images (DTI) A number of 22 patients diagnosed clinically with PD and 25 control cases underwent DTI imaging (TR/TE 4300/90; 12 directions; 4 averages; 4/0 mm sections; 1.2 x

10

Chapter 2. Database and its importance

Figure 2.1: Database input

1.2 mm in-plane resolution) after giving informed consent. This represents, as far as we know, one of the biggest cohort of PD patients implicated in a study. The heterogeneity of the patients - Asians, Eurasians and Europeans - can also be used to characterize a general trend for PD prognosis. As shown in Figure 2.1 the characteristics of the images taken into account for the study present some limitations for certain patients. From this point of view, the image stacks taken for the volume analysis can be placed higher or lower on the body of the subject, hence we have some incomplete studies. Also the image resolution does not permit an automatic detection in some of the cases as it is not able to detect a signicant dierence in the contours of certain anatomical regions. Working with the DTI technique we rely on the fact that water diusion is highly anisotropic in white matter with and the reason for that is that the water molecules were restricted in the axons. The DTI images that we are using were taken with a Siemens Avanto 1.5T (B=800, 12 diusion directions). All the images are in DICOM format. This format is specic to the medical images, containing the header le and the image encapsulated in the "dcm" le. On the header le all the information regarding the patient is contained together with the technical information regarding the parameters for the acquisition. Reading this le we are able to establish the order of the slices and to create the volume image using the same type of DTI. There are several types of DTI images that dier among them by the fact that the coil that takes the images is dierently placed and/or the diusion is performed in several specic directions.

2.2. Echo-planar images (EPI)

11

2.2 Echo-planar images (EPI) From the DTI images the EPI (Fig. 2.2(a)) are among the ones with the lowest resolution. The advantage of this type of DTI is that they contain the tensor information as matrixes that give the actual orientation of the water ow that denes the brain bers. The diusion directions have each of them as result one volume of images. In this way we use for each volume 351 EPI images with 12 diusion direction and the one without diusion and 27 slices. This is the reason why the tensor computation taking the 12 directions into account, has a good accuracy coming from many images.

2.2.1

Tensors and their purpose

The tensors are obtained as a result of the water diusion on the neuronal bers and they are stored as matrix with direction of diusion. This information is able to give as the direction of diusion, as well as the anisotropy values stored as tensor values. To make use of this information we limit the value of the anisotropy for noise. The tensors are computed using the diusion directions and the B0 image as ground truth. Represented as directional-related indices, the tensors oer information regarding the angulation between the current location of a ber and the possible evolution of the same ber. This angulation is limited in our case at 60 degrees minimum, to avoid noise.

2.2.2

Managing the information coming from the EPI

This type of image is not tting for the anatomy extraction and analysis, but the tensor and anisotropy values stored represent the bottom line of the ber reconstruction, as well as the source for other images. We perform the entire image preprocessing on the EPIs, as we need the best values for the bers. A preprocessing step for these images represent a contrast enhancement of 0.5 % for a better detection of the skull and the volumes of interest. On this image we perform the skull removal for the detection of the center of mass and the edge detection on the brain for the detection of the starting point for the inter-hemisphere axis. The removal of the skull is needed because the results presented in the previous report have shown inuences of the voxels that are not brain tissue on the anisotropy analysis. We perform the removal task on this type of images because they contain the anisotropy values and the tensors that are used further in the detection of the bers and they must be uninuenced by any voxel intensity other than the brain tissue.

2.3 Fractional Anisotropy (FA) images Fractional anisotropy images are the result of the computation of the anisotropy level for each voxel on the EPIs(Fig. 2.2(b)). They contain not only the anisotropy

12

Chapter 2. Database and its importance

(a) EPI example

(b) FA example

Figure 2.2: EPI and FA example images values, but also the color code for this type of image represents the diusion direction inside the bers. Because of that, the putamen area is well dened as it is surrounded by the motor tract and stands out as contour with high anatomical detail so we use it in the automatic detection of this volume of interest. After a registration of the obtained volume of interest extracted from this image we can use it together with the tensors from the EPI to limit the bers that we take into account. At this point there is an interchanging of information from one image type to the other one, by information fusion.

2.4 Other types of DTI images and their importance The medical doctors swap between several image types when they are studying a patient because each image has a speciality. In this process the EPI is not used for the visual diagnosis. Usually the T1 and T2 images are used for this purpose because of their high resolution (Fig. 2.3). This aspect persuaded us to consider one of these images for extracting the volumes of interest, especially for the putamen area, which is not well delimited usually. We actually tried to extract this volume form the T1 image (see gure 2.3(a)), but the limitation between the putamen and globus palladi is not clear and the lack of accuracy in this matter can determine us to take together with the bers from the motor tract, other bers as well. After choosing the appropriate image type form the anatomical point of view, before any detection algorithms are applied and tested, we need an image as good as possible, with good anatomical detail and the tensor information for a ber analysis to be possible. This chapter makes the link between the DICOM image detection and preprocessing step with the high level processing and analysis algorithms applied on an image that has better quality. We are eliminating at this point the artifacts due to the specic image acquisition elements and the skull, for it interferes with the detection and processing steps.

2.5. Preparing the image for processing

(a) T1 image example

13

(b) T2 image example

Figure 2.3: T1 and T2 example images

2.5 Preparing the image for processing Due to the complex structure of the medical image encoding manner for the DICOM format we need to take the useful information from the header le and during the processing and analysis steps we only make use of the image by itself, without the additional information. This is the reason why we transform the image from the DICOM format to Analyze and store it as stacks of images that represent an entire brain volume for each patient and each modality. The Analyze format is similar to the DICOM one, except that the header le and the image le are stored separately. Also the header le does not contain so many information as the DICOM one does it does not have any information regarding the acquisition method and parameters (e.g. angulation for the acquisition plane, the series type for the image, the slice number, the diusion direction). When placing the images on a stack, the alignment in between the slices is highly needed as it provides a clean volume image providing a clean contour for the anatomical volumes. For the axial plane the images that we have in our database are taken in AC/PC plane - Anterior Commissure/Posterior Commissure.This line is signicant from the anatomical point of view and it is used by the radiologist because it is distinguishable in all the MRI images. The sagittal plane and the coronal one are not used by our approach.

14

Chapter 2. Database and its importance

2.6 Conclusion Our system of reference in analyzing the results and elements indicative of the PD is represented by the H&Y scale. This scale stands for the diseases degree, 1 being the mildest and 5 the most severe. Unfortunately, for several patients there is no clear classication (1.5 or 2.5 on H&Y scale) and for those cases our system would be a conrmation for placing the patients on the 1 or 2 class for sure, but at this point, our system might put these patients on either of the classes with no mistakes. The preprocessing step performed on these images represents a system able to take from all the DICOM images the ones specied and place them on a volume stack. The mild processing for the EPI and FA images is meant to enhance the success of the automatic detection of the volumes of interest and prepare the images for that purpose.

Chapter 3 Our Approach on Image Processing and Analysis

Contents 3.1 Overview on the system . . . . . . . . . . . . . . . . . . . . . 15 3.1.1 3.1.2 3.1.3

Comparison with other type of images and correlation current diagnoses system . . . . . . . . . . . . . . . . . Proposed approach . . . . . . . . . . . . . . . . . . . . . PDFibAtl@s : our system for diagnoses . . . . . . . . .

with . . . . . . . . .

16 17 18

3.4.1

Automatic detection of the volume of interest (VOI) . . . . .

23

3.2 Removing the artifacts & the skull . . . . . . . . . . . . . . . 20 3.3 Preliminary testing. Motivation for further study. . . . . . 20 3.4 Detecting the Volumes of Interest (VOIs) . . . . . . . . . . . 22 3.5 3.6 3.7 3.8

Registration . . . . The Fusion factor Growing bers . . Conclusion . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

26 28 28 29

After presenting the type of images that we are using, we focus on the image processing part in this chapter, the extraction of the volumes of interest and the way we choose the place where we want to extract it from. We take a look rst at the challenge that we are taking on at this point and then after an overview on the tasks that lie ahead in this part, the algorithms are presented as solutions to these problems. This chapter represents my contribution to the diagnosis system, as it contains the presentation of the idea of image-based diagnosis, a new perspective and approach in this area. The algorithms include our home made automatic detection of the slice of interest and the volume extraction, as well as the new measures introduced for the bers.

3.1 Overview on the system Testing several systems that deal with the specic treatment of DTI images we construct our approach based on the needs of our system, as well as on the results

16

Chapter 3. Our Approach on Image Processing and Analysis

obtained from other systems. After taking a look on the existing systems and their performances on our database, we are dening our goals with respect to the other systems, our perspectives and the way we want to manage to vanquish the challenges.

3.1.1

Comparison with other type of images and correlation with current diagnoses system

As the goal for us is to see how our images perform on dierent software platforms we teste the database on several medical softwares evaluating :

• Registration of images - we want to fuse information from several imaging types and for that purpose we will need the registration • ROI/VOI detection - automatic or manual for restricting the number of bers and having the possibility to dene a bundle of interest • Fiber growth algorithm - for managing the information and analyze the motor tract We are evaluating the computation time, as well as the results obtained for the registration, ROI/VOI detection speed and accuracy and the results obtained at the ber level. The dataow inside the system is important for a database as ours. From the INRIA laboratories we test at rst a free collection of dedicated medical image processing software that provides a DTI registration available, as well as a strong ber growth algorithm. The registration method seams a strong one, as it denes dieomorphic demons. This new method for registration is based on Thirion's demons registration [Modersitzki 2004] and is modied by Vercaunteren in [Vercaunteren 2008a] and integrated in MedINRIA Fusion package. This approach is able to make an automatic registration for medical images [Vercaunteren 2008b]. Using the MedINRIA 1 system there is the need to preprocess each patient and to make a registration between the images that we are using. There is no inconvenience except the fact that the registration does not perform with the accuracy that we need it to perform for us to be able to extract the needed bers afterwards. Also the fact that we cannot limit using two volumes of interest the chosen bers, makes us regard another option altogether. The technical reasons also include the manual registration as the best of them, but even like that, the fact that we cannot choose with the volume the bundle of interest makes us renounce at the idea of using this system. We have tested our images on 3DSlicer, a multi-platform free open source software (FOSS) that oers the possibility to compute the bers and visualize the tensors as well. Testing Slicer 2 we are oered with the possibility to choose two volumes of interest for limiting the bers. The registration for the images is done manually by choosing interest points. This time the system is not able to complete 1 2

MedINRIA - http://www-sop.inria.fr/asclepios/software/MedINRIA/ 3DSlicer - http://www.slicer.org/

3.1. Overview on the system

17

the bers due to the great number of images loaded that make it stop before the bers are computed. In this case we cannot evaluate the ber system as it overloads the memory and it is slow in computation. As Matlab has been one of the main environments for medical image processing, we tested the typical software used for DTI images. The SPM3 has been the main tool for a quick view on what we can extract from our images. The tests presented in the previous PhD report are performed using this tool. We are looking from a dierent perspective now at this tool, as we are focussing on the extraction part. SPM5 has provided a good modality to extract and process the medical images, but the Talairach atlas for the whole brain needs interaction from the user for specication regarding the origin of the patient(e.g. European, Asian, etc.). Also at the skull removal algorithm applied on the images that we have, the results were folded images. Typically the brain atlases are used for taking the volumes of interest into account. This is not accurate due to the fact that we have a brain database from Singapore that contains not only caucasians, but also asians so that a mapping with the brain has small chances of being correct. Testing the SPM tool in Matlab we obtain results only on the entire brain analysis and due to the image quality, the skull extraction cannot be properly performed and thus we have interferences with the results on the anisotropy. A specic atlas that contains automatically detected anatomical volumes represents a tool that can be applied to any type of patient. From the same set of Matlab tools we have also tested FSL 4 for better management of the DTI data, but we did not manage to obtain relevant results with our data as the brain analysis and mapping tool had errors at the segmentation level.

3.1.2

Proposed approach

From our preliminary studies we have seen that the motor bers are relevant for the progression of PD. By building a specic atlas for PD, we intent to segment, visualize and quantify VOIs and motor ber tracts. Substantia Nigra and Putamen represent the two VOIs where the motor bers are placed and from where we extract the bundle of interest. These VOIs are the landmarks for an geometry-based registration [Zitova 2003], used to align the volumes for FA and EPI. We have chosen this registration as we are dealing with intra-patient matching and we need a rigid-body transformation as well. The registration provides a common space for the tensors and the anatomical elements. The registered volumes are fused, by acquiring the anatomical part from the FA and EPI images and considering the anisotropy information from the EPI, which contains the tensors and their direction. The fusion enhances the anatomical importance from the color image (FA) taking into account the anisotropy level and the same area from the low resolution one (EPI) as well. However, the anatomical details are taken primarily from the FA image. On the registered volume of interest on the EPI we perform ber growth between the two volumes of interest and we extract the ber density (FD) and the ber volume (FV). 3 4

SPM-http://www.fil.ion.ucl.ac.uk/spm/software/spm5/ FreeSurfer http://www.fmrib.ox.ac.uk/fsl/

18

Chapter 3. Our Approach on Image Processing and Analysis

This approach diers from other fusion techniques, since we are fusing the images at the information level, using the registration for spatial alignment on the brain volume. We do not rely on prior atlases created for structural MRI and no manual intervention is needed in the process[Lehericyr 2004][Vaillancourt 2009]. Also we do not perform prior segmentation on white matter (WM) and grey matter (GM) or use any maps for alignment and registration purposes[Deisseroth 2009][Woodward 2009]. From the fusion point of view we have a two direction fusion of information: from EPI to FA and on the other way as well. We must rst take into account the specic characteristics of each image type. The FA image represents the fractional anisotropy on the dopamine ow inside the whole brain, color coding on this image indicating the direction of diusion as presented in chapter 1. The EPIs have the tensor matrices that provide the direction of the bers for the ber growth algorithm. From this standpoint, the color analysis approach is better on FA image and for the ber growth we must consider the EPIs. However, on the FA image the detection of the midbrain area is not possible since the algorithm takes a part of the CerebroSpinal Fluid (CSF) as well, and this interferes with our further analysis. Hence we must work on the midbrain detection on the EPIs. On the other hand, there are problems on detecting the putamen area even on the T1/T2 images for the reason that there are not clear limits between this area and the ones surrounding it. On the FA image, due to the dopamine enhanced by color coding we are able to make a distinct limitation for the putamen. We use both regions of interest for each step of the analysis part.

3.1.3

PDFibAtl@s : our system for diagnoses

Before detecting the volume of interest we must perform several steps for preprocessing the image, due to the low resolution of the EPI and the problems presented at the beginning of the section: brain size dierences, anatomical dierence based on particularities specic to each individual as well as elements of brain atrophy from aging issue. As shown in gure 3.1 there are several aspects that we want to solve by introducing algorithms that are able to manage the problems found in other systems typically for our images, with applicability to PD. The main bids are the result of the quality of the image as well as the elements linked to the processing part and the analysis of the 3D images. The system introduced here takes the DICOM le and manages the information to extract the relevant features that we are using to construct the 3D image. The demographic elements that inuence the analysis part are extracted at this level as well. EPIs are used for automatic detection of the midbrain area, algorithm that I developed starting from the geometrical placement of the specic anatomy structures in the brain volume[Kretschmann 2003]. Before applying this algorithm a preprocessing stage is needed and the algorithms used at this stage are developed also specic for the needs of PDFibAtl@s and are made by us. Furthermore using the second type of medical image, the FA, makes possible the accurate detection of the putamen area and is specic to our system. In this

3.1. Overview on the system

19

Figure 3.1: Layout of PDFibAtl@s case we need the two images aligned as we use the automatically detected volume of the putamen on the EPI image, making a registration. The registration itself, as well as the volume detection algorithm are original ideas and represent my contribution to this system. Using the algorithm developed by Basser [Basser 2000] we grow the bers using the tensor information from the EPI images as shown by Bihan in [Bihan 2001], but we limit the results to the volumes detected. In this manner we are able to choose just the bundle of interest that represents the motor tract and evaluate the bers using our own metrics. The preprocessing part has to overcome the low resolution of the EPI, as well as the demographic characteristics of the patients (age and sex dierences). In our study, we surmount the sex dierences by computing the volume of each brain, as there is a dierence between the female volume of the brain and the male volume, based on the fact that women usually have smaller skull. At this point we compute measures based on the density of the bers in the entire volume of the brain or of the volume of interest. FN r FN r FD = ; F Drel = (3.1) V olBrain V olV OI where F D represents the ber density computed as the number of bers - FN r in the volume of the entire brain - V olBrain and F Drel represents the ber density relative to the volume of interest- V olV OI . We try to overcome the age dierence as well by taking the mean age on the testing batch as close as possible between the PD patients and the control cases.

F V = FN r ∗ Vheight ∗ Vwidth ∗ Vdepth ∗ Fleng

(3.2)

where F V represents the ber volume computed as the product of ber number (FN r ), ber length (Fleng ) - constant as the bers must pass through both regions of interest and the voxel dimensions: Vwidth , Vheight , Vdepth . According to the medical manifestation of the disease, the ber density and volume should be diminished for the PD patients compared with the control cases and the degradation of the bers should be correlated with the severity of the disease specied by the H&Y scale.

20

Chapter 3. Our Approach on Image Processing and Analysis

For detecting the elements related to the volume of interest we take the relative position of anatomical elements to a x point. We have chosen this point to be the center of mass of the brain (Xc , Yc , Zc ). In order to determine this point we need to have the brain, without the skull. Also we need to compute the brain volume to determine on which slice to look for the volumes of interest.

3.2 Removing the artifacts & the skull The process of removing the artifacts is based on the fact that all the needed information resides inside the contour of the skull. These limits are not well detected, as the artifacts follow the shape of the skull, just like an aura. The open/close operation that can be performed on these images have eect on the brain tissue as well, because the intensity of the pixels have the same value as the one of the artefact. In this case we take as exterior limit for our algorithm the skull and we remove in the same time the skull and the noise from the image. For this purpose we use a segmentation method that is not so sensitive to the intensity of the voxes and classies the information into "tissue" and "bone". This method does not detect the noise, but only the useful information and the noise and artifacts that were confounded with the brain tissue, primarily with the CSF, are placed in the "tissue" class. Using the KMeans algorithm implemented as plugin in imageJ takes as input the number of classes that we want to achieve at the end. The mask obtained for the brain has two purposes: to act as an exterior limit for removing all the voxels outside of it and to remove the skull that is represented in the mask that we Figure 3.2: Image of the brain contour are using. This preprocessing allows us to have only the necessary information to compute not only the anisotropy, but also the geometry elements for the volume and the bres, as well as the starting points for the volume automatic detection.

3.3 Preliminary testing. Motivation for further study. Our preliminary work is based on the determination of the rst region of interest volume of interest where a study is performed on the FA images considering the color

3.3. Preliminary testing. Motivation for further study.

21

Figure 3.3: 3D view of EPI stack with B0 images coding specic for this type of images. The colors for this type of image represent the direction of diusion into the ber tracts:

• red - left right (LR); • green - anterior posterior (AP); • blue - up down(UD); Based on the fact that the motor tract should go in anterior -posterior direction we make a color analysis on the volume of the midbrain, extracted in order to see whether the bers starting from this area have a correlation with the H&Y scale. Figure 3.4 represents the main steps in the analysis of the green channel. We make the detection of the midbrain area and compose the volume of interest on the EPI image and then place the determined volume on the FA image for the green channel extraction, performing an alignment between the two image types. Once we split the volume obtained from the FA image on the three channels and take the green one, we perform the histogram and extract the values for the range of interest. This range is chosen in a way that we can exclude the noise and we place it in between 10 and 100. These values are then correlated using PASW 18.0 (Predictive Analytics SoftWare, formerly SPSS) tool with the H&Y values. An analysis of the methods used for correlation, as well as the testing procedure used is presented in chapter 4. After this study we continue with the growing of the bers to examine their density and the relation at a higher granularity level with the disease.

22

Chapter 3. Our Approach on Image Processing and Analysis

Figure 3.4: Green channel analysis This analysis oers the opportunity to see if bers starting from the midbrain area, where the dopamine is produced, are aected by the PD and the correlation between the disease and the level of green. This level of green represents the anisotropy level as well and this particular area is the one that is correlated with PD, as most of the bers going in AP direction are represented by the motor tract. In this case the anisotropy level, if correlated with the H&Y scale represents an indicator of the disease at the midbrain level. Therefore we are able to determine that the starting point for the bers we want to grow is relevant for our study.

3.4 Detecting the Volumes of Interest (VOIs) Knowing that the motor tract starts from the midbrain area and the anisotropy level is correlated with the PD severity we take the bers from the motor tract that cause the trembling for the PD patients and study their correlation with the disease. The volumes of interest limit the bers that we grow and in this manner we are able to select only the ones that constitute the motor tract. That is why a good denition for the volumes determines a good bundle of interest for analysis and aects the nal result of our study. For that purpose, as the midbrain area contains many bers appertaining to other tracts than the motor one, we need another factor to choose from these bers. The putamen area is another anatomical location where this tract is passing through and should contain all the bers from this tract. In this case we take the putamen area as the destination of the bers that we grow. Choosing for analysis only the bers that cross in the same time both volumes of interest, we select the bundle of interest for our study. For reaching our goal: extracting the bers that constitute the motor tract, we need the volumes of interest, but we need to extract these volumes from the images

3.4. Detecting the Volumes of Interest (VOIs)

23

that we have. The algorithms for extraction must be placed at the right place in the 3D image volume, for this detection to be as accurate as possible. Placing the VOI detection algorithms is possible once the slice of interest(SOI) is detected. Detecting the slice of interest starting from the center of mass of the brain is done by taking into account the placement of the anatomical regions that we consider as volumes of interest. For the midbrain we consider the slice of interest 8 mm lower than the center of mass and for the putamen area 2 slices higher than the center of mass, thus 8 mm higher than the slice containing the center of mass as well. Due to this manner of placing the slice of interest according to the center of mass there are several patients that do not perform well. These are the patients that in the volume of interest do not have the entire brain and the volume is shifted towards the neck more than the brain. In this way the patients do not have all the slices containing all the upper skull and the brain.

3.4.1

Automatic detection of the volume of interest (VOI)

Once we have the slice of interest detected for each of the volumes used for the bers, we need algorithms that determine the placement in the image slice of the region that we want to detect and then take all the voxels that are part of the region of interest. Knowing where in the image we have the regions, we do specic algorithms for each volume for the stating point and the detection.

Detection for the starting point of the volume of interest in the midbrain area is done similar to the detection of the slice of interest and it is combined

with the division in hemispheres of the brain. We need the hemispheres separately on account of the study of Dr. Chan [Chan 2007] which states that there are different stages of development of PD in the left side of the brain and the right side. Usually the left side of the brain has more bers grown in between the two chosen volumes of interest. The actual algorithm that makes the division into hemispheres for the brain takes the contour of the brain in the slice of interest. Based on the inexion point with the maximal value we mark this point as a part of the medial line and together with the center of mass detected in the slice of interest the two points represent the medial limit in between the two hemispheres of the brain. This limit is used when we detect the volumes of interest as we want the algorithm to take only the needed area for the volume of interest.

In the algorithm for detecting the volume of interest in the midbrain area we have two steps for detection: the denition and detection of the region of

interest and the volume detection. For the region of interest we use a snake-based algorithm applied on a segmented image with KMeans in imageJ. We segment the image in imageJ for we want to make the dierence between the Cerebro Spinal Fluid (CSF) surrounding the midbrain and the area we want to detect. As in the image we have white matter (WM) tissue, gray matter (GM) tissue, CSF, as well as bone from the skull, elements that were not eliminated by the algorithm, together with noise and artifacts - pixels surrounding the brain as an aura due to the quality of the image, a classication algorithm based on the intensity of the pixels is needed.

24

Chapter 3. Our Approach on Image Processing and Analysis

(a) EPI with detected mid- (b) EPI with detected puta- (c) 3D image of bers debrain and bers men and bers tected passing through VOIs

Figure 3.5: EPI with detected VOIs in 3.5(a), 3.5(b) and 3.5(c) with 3D bers on an example We use these types of tissues and elements as classes for the KMeans algorithm in imageJ. On the gray matter class we perform the snake-based algorithm that has the starting point as the center of mass in the slice of interest and depending on the side of the brain that we want to explore, our algorithm selects each pixel and compares it with the anterior pixel. This exploration step ends when there is a dierence between the new pixel and the previous one or we step on the midline of the brain. After nishing the algorithm on one slice we explore the slice above in similar manner. As we know from the study presented in [Starr 2009], almost 80% of the SN is found in one slice (4 mm) thus we want to make sure that in our volume of interest this anatomical region is contained and for this purpose we take the two slices that most probably contain this region. In gure 3.5(a) there is the detected volume of the midbrain for both sides of the brain hemispheres and the bers projected on the EPI. In gure 3.5(b) , after detecting volume of interest on the FA images and registering it on the EPI for both sides we use it for choosing the bers. On the image 3.5(c) we have a 3D view of the bers passing through the detected volumes of interest. The algorithm for detecting and growing the bers is the one presented in [Westin 2002]. This algorithm is made for the WM bers that are grown in the same direction, unlike the ones in the GM. The VOI containing SN is GM tissue hence we must adapt our algorithm for our purpose, to have better bers and clearer result.

Detecting the starting point for the putamen detection algorithm

is dierent from the one used for the midbrain, as the putamen is not placed on the midline and does not have a geometrically detectable point or standard distance. That is the reason why we base our algorithm on the intensity values of the voxels, but is based on the placement of the two areas relatively to the center of mass of the image. The algorithm starts from this point and following the midline it stops at half of the distance between the center of mass and the upper limit of the brain area. At this point the intensity of the voxels on the left and right side of the midline are

3.4. Detecting the Volumes of Interest (VOIs)

25

parsed and after passing through the globus palladi area with higher intensity than the midline, once we reach an area with low intensity, the detection algorithm can be performed.

For the Putamen volume detection we take into account the shape of

this specic anatomical region: triangular. As the FA image is more clear than the other type of images due to the dopamine ow, we perform this detection on this image. The slice of interest for the detection algorithm for this area is the one where the center of mass of the brain resides. Starting from the axis that separates the two hemispheres, the center of mass is placed on this axis and at half the distance from this point towards the frontal lobe we are placing Figure 3.6: The main anatomical struc- the seeds for the volume detection on tures at the putamen level of the brain the left and right side of the axis. The [Talos 2003] placement relative to the axis is chosen by taking into consideration the intensity of the globus pallidus much lighter than our region. Once we cross that area we encounter the putamen and we place the seeds for our algorithm. This algorithm is represented by triangular formation that stretches on the area that has a similar intensity with the seeding. The area determined in this manner has a triangular shape, as we have three seeds that are moved by adding new points inside the triangular area. The positions of the three initial seeds are the nal limits of the regions of interest. We extract the area inside the triangle and we register it on the EPI image. Comparing with the traditional atlases like the Tailiach atlas, our approach is intuitive and is based only on the image information doubled by the anatomy map that represents the geometrical disposition of the main brain structures. When testing the existing atlases we had problems with our images because the database is heterogenous and the mapping of the brains for anatomical extraction of the specic brain structures was not successful. Either the structures were not correctly identied, or we needed to identify other structures as well. The SPM could not place the brain atlas on the image because we did not specify the demographic provenance of the patient. The PDFibAtl@s has the geometrical element that is able to detect the slice of interest for the volumes, but also the intuitive algorithm based on the intensity of the voxels for the volume extraction. In this case we do not need the demographic information, as we start the algorithm by detecting the region of interest where we grow the volume. This approach is considered an atlas, as we take into account the placement of our volume of interest relative to the center of mass of the brain and the other anatomical structures that are placed around our

26

Chapter 3. Our Approach on Image Processing and Analysis

volume.

3.5 Registration Talking about registration we refer to matching or bringing the modalities to spatial alignment by nding the optimal geometrical transformation between corresponding image data. The challenges for performing the registration reside in nding the best landmarks in both image types nding a suitable spatial transformation and for our type of images preserving the tensor direction. For our case we perform intrasubject registration, as we are matching images appertaining to the same subject; our registration is a rigid one, as it contains only translations and rotations and fully automatic. It is a rigid body transformation because we do not change the shape of the elements. Also due to the fact that we are using homologous features that are mapped using geometrical distances our registration is a geometrical-based one. For the midbrain area we use the EPI, as it is clear enough for this purpose, even if the resolution for this type of image is poor. We cannot do the same thing for the putamen area and even on the high resolution images like T1 and T2 the margins of this anatomical region are not well detected by the algorithms. In this case we use the FA image and take advantage of the anisotropy dierence presented in this type of image as intensity dierence. This makes possible a detection on the FA image of the putamen. However when we use the detected image we want to do that on the EPI image and we need to know that the extracted volume is on the right place. For this purpose we verify that the placement of the volume of interest is relative to the center of mass of the brain, as well as the external limits of this volume are related to the same point. In order to determine the directionality of the image we use the symmetry axis and its orientation. It gives us the angle with the horizontal and vertical for the rotation and the displacement. All the transformations are performed on the image extracted from the FA image, keeping the EPI. Analyzing the technique we used we can say that we perform an iconic registration [Cachier 2003] because we use on one hand the geometrical relations and placement of the center of mass and the external limits, but on the other hand we use the anisotropy values for dening the volume that we register [Elsevier 2009]. As we are not using that information directly for the transformation of the image, our registration is more geometrical [Golinpour 2007][Maintz 2000].





0

x

y

0

z

0

cosθx sinθx  −sinθy cosθy 1 =  0 0 0 0

  0 dx x   0 dy  y   1 dz   z  0 1 1

(3.3)

Representing the transformation applied on the FA image in equation 3.3 we represent the rotation, translation and skewness. The rotation angle for the transformation is computed by taking into account the symmetry axis determined for

3.5. Registration

27

Figure 3.7: Geometrical view of the registration parameters delimitation of the two brain hemispheres. The θx value is the angle between the axis and the Ox axis of the image and the θy is the angle between the same axis and the Oy of the image. We compute this angle for each image type and the dierence between these angles represents the values for the transformation.

SPy I1 SP SPx sinαy = I2 SP

sinαx =

(3.4) (3.5)

where SP is the starting point of the hemisphere axis given by the inexion point placed on the lower part of the brain (posterior area of the brain) and the SPx and SPy are the projections of the SP point on the Ox respectively Oy axis; I1 is the intersection between the axis and Ox and I2 is the intersection between the axis and Oy . We compute the α angle for the FA image and the β angle for the EPI image. The θ angle is the dierence between α and β and we use it for the rotation. The translation valued from the transformation matrix from equation 3.3 (dx , dy and dz ) represent the dierence of the center of mass in the two type of images. Another aspect of the transformation is represented by the axis orientation. The dierence between the orientation of the axis determines us to ip the transformed image. This orientation is determined by the placement of the starting point (SP) and the center of mass on the image axes. Dierent orientation of the axis determines a ipping of the image in horizontal and/or vertical plane. We determine the orientation of the axis by computing mathematically the equation of the line

28

Chapter 3. Our Approach on Image Processing and Analysis

determined by two points : SP and the center of mass. The position of the SP point on one image is the inexion point placed on the contour of the brain. The orientation of the axis is determined by the projection of the points on the image system of references. Because the FA images are generated using the EPIs, there could not be any skewness problems or resizing aspects, thus we are performing the transformation only for the translation and/or the rotation aspects. As the FA images have dierent orientation we need to be sure that the volume of interest is correctly placed on the still image.

3.6 The Fusion factor Fusing two images means morphing them or warping them. Both these techniques represent registration methods used to alterate one of the image by incorporating the information from the other image. Making the registration we alter the extracted volume of interest by adding the orientation of the EPI and its geometrical characteristics. This can be considerate as a fusion at the image level, but in our case looking at the bigger picture we realize that we transfer knowledge detected on the FA onto the EPI. In this case we are talking about fusion from other point of view as we do not want to change the image. From our point of view we are performing the fusion by taking the medical knowledge based on the atlas-mapped positioning of the detected volumes, but also by using the obtained volume as a mask for limitation reasons on another image type. Putting together information from dierent sources enhances common characteristics and adds the specic elements from each sources. We fuse the information in this case by taking the anatomy for the putamen from the FA image and the tensor information in the EPI. We fuse the two images without blending them [Zitova 2003] or warping them [Golinpour 2007], just taking the needed information from one image and inserting it into the other one by using the registration [Maintz 2000][Wirijadi 2001]. The added value of knowledge from both imaging types prepares the landscape for the bers growth algorithm.

3.7 Growing bers Before we talk about the modality to detect and select the neural bers, a denition for this concept is necessary from the medical point of view. From the anatomical point of view the gray matter(GM) tissue is made by the dendrites of the neuron and the white matter(WM) is made of the axon of the neuron. Neural bers represent the linkage between the axon of a neuron with the dendrite of another neuron. The anisotropy represents the enhancement of the neural ow passing through the axons. The midbrain area, where the SN resides, is a gray matter volume. The

3.8. Conclusion

29

process of growing the bers starting from the EPI means actually taking the tensor information and based on the anisotropy value choosing the starting point of the bers. In the white matter area, the placement of the bers is more obvious because the the axons represent this area and the neural ow is very intense. That is the reason why for us is very challenging to do the ber recognition and growing starting from the midbrain area, where the predominance of the tissue is the GM. At rst we implement a classical ber growing algorithm based on the white matter(WM) area because like that we can compare our algorithm on the same set of images with an existing one and we can verify the position of the bers in the volumes of interest. For our system we take into account the approach presented by Basser in [Basser 2000] and for the tensors approach we use Bihan's approach [Bihan 2001]. In Basser's approach the algorithm is based on the Fernet's equation for the description of the evolution of a ber tract. This approach is specic to white matter, as the axons are the white matter. The midbrain area is gray matter. Growing bers from the gray matter is a challenge since the number of axons in this area is much less than in the white matter and the bers are not so well aligned as the ones in the white matter. The tensors represent oriented segments that are placed on consecutive layers inside the volume. Starting a ber means taking the tensors and computing the match of two consecutive tensors. If the value correlated with a lower threshold for the anisotropy of 0.3 - specic for the algorithm, and an angle for the ber that is not higher than 60 -obtained from the rellevance values of the bers[Chan 2007]. We rst apply this algorithm in order to see if there are relevant bers that we can grow between the two VOIs ( Fig. 3.5(c)). We use these VOIs to choose the bundle of interest and separate the bers that we need from the ones that are not part of the motor tract. Although we grow all the bers, we validate only the ones starting from the midbrain area that reach also the putamen area. Fibers too small or those that do not go towards the putamen area are not validated. In this manner with the second region of interest we have an element that validates the grown bers, without needing the SN clearly dened.

3.8 Conclusion The green channel analysis is done based on a similar variation on the histogram for the volume of interest on this channel, indicating that the bers all go in AP direction and that there is a correlation between the anisotropic green channel and the H&Y values. Furthermore the bers that start from that area and grow towards the putamen have an even stronger correlation with the PD evolutive scale. At this point we are able to automatically detect the volumes of interest by selecting the slice of interest where we are exploring this volume. Determining automatically the limit between the two hemisphere allows us to make a dierentiate analysis for each side of brain. We need this possibility as the PD condition has

30

Chapter 3. Our Approach on Image Processing and Analysis

dierent level on each side of the brain for the same patient. That is the purpose for separating the left and right side of the brain for the ber growth algorithm. The registration combined with the fusion factor is specic for our imaging types, as well as for the anatomy in question. This is based on the geometry registration enhanced by the brain anatomy elements and this algorithm eliminates the specicity elements of the processed case.

Chapter 4 Validation, Results and Interpretation

Contents 4.1 Medical relevance . . . . . . . . . . . . . . . . . . . . . . . . . 32 4.1.1 4.1.2

Green Channel analysis on the midbrain area . . . . . . . . . Fiber study . . . . . . . . . . . . . . . . . . . . . . . . . . . .

32 33

4.2.1 4.2.2

Medical Image processing . . . . . . . . . . . . . . . . . . . . Speed for computation . . . . . . . . . . . . . . . . . . . . . .

36 37

4.2 Technical specic elements . . . . . . . . . . . . . . . . . . . . 36 4.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Validating our methods is as important as the algorithms themselves. This chapter treats the manner in which we validate each of the new algorithms, as well as the results obtained. The validation can be done by comparing the results obtained with other softwares/algorithms on our database with our results. Another level of validation is represented by dissemination. There are several stages where we evaluate our system in order to see at each point the relevance and its impact on the whole program. As presented in chap.3, one rst stage for evaluation is after the automatic detection of the midbrain volume of interest where we perform an analysis of the green channel from the FA image. The fusion is done by detecting the VOI on the EPI image automatically and registering it then on the FA image for the color analysis. The second stage is represented by the ber analysis and the fusion of information is done by detecting the second region of interest on the FA image and registering it on the EPI for the ber limitation. We have on each stage a validation and an evaluation stage for the proposed method. On the rst stage we take into account the heterogeneity of the subjects from the demographic point of view and we analyze the eects of each element in the testing area in order to nd a test that is not sensitive to these elements and meaningful as result. From the initially 47 cases, 42 of them have been successfully processed by our system and the correlation and analysis of the results is based on these cases. The cases that could not be processed have either corrupted images or the slices have been taken too low hence the automatic detection step cannot be initialized.

32

Chapter 4. Validation, Results and Interpretation

4.1 Medical relevance For the medical relevance we use the T-Test for correlations between the obtained values and the cognitive evaluation as ground truth. We validate the volumes obtained by running the results through our neurologist to verify the placement of the detected elements on the initial images. The automatic detection is validated also using the manual detected volumes and by extracting one of the from the other, the dierence represents the error rate. For the bers we can only verify that the ones chosen are approximately on the SN area. This is also done by our neurologist.

4.1.1 Green Channel analysis on the midbrain area For the green channel study we have a batch of 42 cases (21 patients and 21 control cases) where we take out randomly 5 cases from the patients and controls in order to eliminate the subFigure 4.1: The Results Window jectivity from our study, the inuence [Teodorescu 2009d] [Teodorescu 2009a] of the demographics in the green channel study (table 4.1). The T-Test is applied on the histogram obtained from the midbrain area by eliminating the noise. We study in this manner to see whether there is a correlation between the value of the histograms and the H&Y values. The histograms represent the anisotropy value on the AP direction in the midbrain area, which should indicate the motor bers and in PD could be characteristic for the progression of the disease. Examining this correlation we variate the age dierence between the patients and controls and the number of male subjects in the testing batch, as well as the mean value on the H&Y scale. The results on the green channel study performed on the patients with the characteristics from Table 4.1 are presented in Table 4.2. This table contains several T-Test methods and their results regarding the correlation between the green channel histogram values and the H&Y values. When we analyze the Independent Sample T-Test we have a large variation between the values of P, which can be explained only by the variation of demographic characteristics of the patients. In our case this type of study is aected by demographic characteristics,

4.1. Medical relevance

33

Test

H&Y

Age

Male/all

nr

[avg]

Patients

Controls

Pat

Control

1 2 3 4

2.312 2.375 2.375 2.467

64.5 63.31 64.06 62.75

59.37 60.93 58.5 61.5

11/16 9/16 8/16 9/16

6/16 9/16 7/16 8/16

Table 4.1: Test batches characteristics [Teodorescu 2009b] Independent Sample T-Test

Test

Correlate Bivariate

ANOVA

nr

Left

Right

Left

Right

Left

Right

1 2 3 4

24.4 12.2 75.5 83.6

74.0 69.3 65.3 71.4

13 7 3 7

8 8 6 7

0.872 0.906 0.937 0.937

0.937 1 1 0.906

Table 4.2: Study on Green channel on the left and right side [Teodorescu 2009b] especially on the left side ( e.g 12% - 83 %). We almost have the same range of variation on the Bivariate test, visible on the left side as well. The ANOVA test is the most coherent and has good results. This test is reliable and adequate to our purpose. An initial approach and the associated results have been presented on the RSNA conference [Teodorescu 2009b] from the clinical point of view. We emphasize here the technical asset.

4.1.2

Fiber study

The motor tract is automatically detected in our case by growing the bers between the two volumes of interest: midbrain area and the putamen. After computing the FD and FV on each side of the brain we study the eects of PD in each bundle of interest. For this purpose we perform the T-Test making the correlation between FD/FV and H&Y scale. As the FD is dependent on the FV, the two parameters have the same variation. For the medical relevance on correlating the H&Y parameter with the bers we test the obtained values using WinSPC (Statistical Process control Software). For the simple correlation purpose we analyze Pearson's parameter (see Table 4.3 column 2 and 4). We have chosen for testing in this case the ANOVA method : one way ANOVA, General linear model ANOVA (MANOVA) and we test the equal variation on density considering the Lavene parameter. When we perform the global testing taking into account 80% of our data we obtain p=0.05 for the group homogeneity in the H&Y assigned cases classied using

34

Chapter 4. Validation, Results and Interpretation

Figure 4.2: 3D View of the grown bers the left ber results. On the ANOVA test for the same cases the signicance is 83% with an N=35 subjects randomly taken by the software from the 42. Taking a closer look on the testing batches we can follow the variation of the relevance degree depending on the demographic elements and with regard to the test taken. A signicant value for correlation is given when the value of Pearson variable is lower than 0.01. In table 4.3 we perform the testing for correlation between the H&Y value and the FV. Our conclusion after this test is that it is inuenced by the testing batch that we take into account. For test batch 3 on the left side we have both variables indicating a very strong correlation, while the other test vary and appear not to be signicant. For the results presented for the bers we perform the same way of testing used Test

Left Side

Right Side

nr

Pearson

P-value

Pearson

P-value

T1 T2 T3 T4 Total

0.041 0.107 0.010 0.108 0.054

0.825 0.555 0.955 0.555 0.735

-0.096 0.023 -0.037 -0.101 -0.098

0.599 0.898 0.841 0.581 0.541

Table 4.3: Simple correlation between the FV and H&Y values[Teodorescu 2010]

4.1. Medical relevance

35 One-way ANOVA

Test nr

FV

T1 T2 T3 T4 Total

MANOVA FD

FD

Left

Right

Left

Right

Left

Right

0.00 0.00 0.00 0.00 0.00

0.00 0.00 0.00 0.00 0.00

0.00 0.00 0.00 0.00 0.00

0.00 0.00 0.00 0.00 0.00

0.105 0.638 0.138 0.329 0.149

0.515 0.067 0.404 0.404 0.629

Table 4.4: ANOVA testing [Teodorescu 2010]

Test

Test variation of density P-value

Lavene

nr

Left

Right

Left

Right

T1 T2 T3 T4 Total

0.499 0.932 0.888 0.733 0.742

0.855 0.542

1.33 0.000

0.04 0.57 0.721 0.921

Table 4.5: Variation of density [Teodorescu 2010]

for the green channel analysis. This time we test our batches of patients taking into account the correlation and the regression coecients. We can distinguish a dierence between the two sides of the brain on the results tables. The variations among the testing batches is due to the dierences between the subjects. One -way ANOVA test is used to compare three or more unmatched groups and that is the reason we test our results using this test (rst 4 columns in table 4.4). MANOVA results are presented in columns 6 and 7 from the same table. On the ANOVA Oneway test the value considered signicant is 0.00. In table 4.4 we can conclude that this test shows a strong signicance on all the testing batches, while the MANOVA and the Lavene variable don't show a signicance. In some of the cases the equal variation of density could not be computed due to lack of a certain type of cases (Table 4.5), while the Lavene parameter is signicant only the whole database on the left side. These T-test show the medical relevance of our system, but from the technical point of view we have to evaluate the robustness of the algorithms and their speed, as well as their accuracy compared with the manual detection and extraction.

36

Chapter 4. Validation, Results and Interpretation

4.2 Technical specic elements From the technical point of view the relevance of the system resides in the automatic detection of the volumes of interest and the management of the medical images, as well as the ber growing algorithm. For the medical image processing we deal with the DICOM format. This specic medical format for image consists form a header le and the image information encapsulated in the same DICOM format. The header le contains the patient information, as well as the angulation and the type of the image. We are parsing the images from a folder reading the patient id and the image type and once we have the type that we need (EPI) we read the slice number and the direction of diusion for making the volume for the patient that we are dealing with. All this preliminary steps are performed using imageJ 1 toolbox in Java.

4.2.1

Medical Image processing

We use Java for all the system with imageJ toolbox and bio-medical imaging plugins 2 . The simple image processing for the preprocessing part is done by enhancing the contrast for the EPI images and removing the noise. For the removal of the skull we use K-Means for making the segmentation based on the pixel intensity. By removing the skull we remove the outside noise surrounding the entire brain, the aura eect induced by the scanner. For the 3D visualization we are using the Volume Viewer from imageJ 3 . For computing the inexion point at the bottom (posterior) of the brain we use the automatic contour detection and by analyzing the contour as a variation function we take the inexion point and mark it on the image as the limit between the two hemispheres. This technique was visually validated by constructing the midline or inter-hemisphere axis using the center of mass of the brain as well and the equation of the axis determined by these two points. On the evaluation of the detected volume we compute the dierence between the manually detected volume and the automatic one. This dierence represents the error rate for the algorithm. Also a validation done by our neurologist is necessary for this step. For the registration performed on the detected volume we use the medical knowledge for validation and a visual evaluation. The fact that for the volume detection of for the registration other algorithms do not deliver satisfactory results from the medical point of view, determined us to create our own algorithms, specic for the type of images that we are using. Other algorithms previously used for volumes such as the snake algorithm or the volume VOI automatic detection do not perform well on low-resolution images such as the EPIs. For the registration we tested the TurboReg 4 but as it was not developed for head images, it did not 1

imageJ -http://rsbweb.nih.gov/ij/ Bio-medical image -http://webscreen.ophth.uiowa.edu/bij/ 3 Volume Viewer 3D - http://rsbweb.nih.gov/ij/plugins/volume-viewer.html 4 TurboReg - http://bigwww.ep.ch/thevenaz/turboreg/

2

4.3. Conclusions

37

perform as expected. The registration is very important, as we have seen also testing the MedINRIA tool [Vercaunteren 2008a] [Vercaunteren 2008b] [Modersitzki 2004]. This tool due to the registration did not nd the bers that we needed for the analysis. The 3DSlicer [Ceritoglu 2009] performed good on the manual registration and even like that it was not only time consuming, but also from the resources point of view disappointing as it crashed the machine each time at the end of the process.

4.2.2

Speed for computation

The software was tested on Intel core Quad CPU Q660 (2.4GHz; 4.0G RAM) and the average time for each patient was 4.68 min with the automatic detection and the ber growing algorithm. If with MedINRIA took us 1-5 min to have the bers, with our tool it takes us an average of 2 min. This is because we limit the area for the bers and we do not grow them on all the brain and take the ones we need, but just grow the ones that cors the specic volumes. The 2 mins represent for our system all the processing time with the automatic volume detection, registration and the bers.

4.3 Conclusions From the point of view of the evaluation and resting criteria the contribution reveled on this chapter reside in the testing technique. This technique mixes the cases for revealing the eects of the cognitive parameters and the test that are not aected by these parameters. The green channel analysis has been presented in [Teodorescu 2009b] [Teodorescu 2009d] [Teodorescu 2009a] already. The bers and the PDFibAtl@s approach as an entire system is presented in [Teodorescu 2010] and a s a demo version at [Teodorescu 2009b]. Another important aspect at this point is nding the appropriate test for the data that we are provided with. If for the green channel study we take into account one set of tests, the one that is most appropriate for the histogram, for the bers density, due to the normalization, we cannot take the same test and we need something more appropriate for correlation and regression values. Evaluating the obtained volumes of interest, as well as the techniques implemented, they proved to be appropriated for the type of image that we are dealing with, as well as for the resolution of these images. The speed of computation reveals a system that performs in a few minutes the detection of the regions of interest, as well as the computation of the bers. We have found a way to evaluate our algorithms separately and the whole system as well. The testing approach has been validated by dissemination and the algorithms by comparison and dissemination as well.

Chapter 5 Conclusion. Further study. Applicability

Contents 5.1 5.2 5.3 5.4 5.5

Diagnosis evaluation . . . Prognosis potential . . . PDFibAtl@s importance. Contribution . . . . . . . Conclusion . . . . . . . .

............. ............. Future applicability. ............. .............

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

39 40 40 41 42

Our approach is important from the clinical point of view oering a new tool for the neurologists in PD and a mean to verify/conrm their diagnosis and propose a prognosis. From the technical standpoint, the fusion is novel, as it combines the tensor based information and the anatomical details. This system provides data for H&Y estimation and PD prognosis. The main new breakthrough is represented by a method able to predict the PD, as well as an evaluation method based on the image attributes, on the anatomical and neurological aspects of the patient. As the H&Y test is based on the cognitive facet, our method is complementary to the test, but is placed on the same scale. PDFibAtl@s is a new system that is able to automatically detect the volumes of interest for PD diagnosis using the DTI images and a geometrical approach. The algorithms for this task are home made and are based not only on the brain geometry, but also include the medical knowledge by taking into account the position of dierent anatomical structures at the brain level, hence the atlas dimension. As for the fusion contribution of our work, it brings together the FA clarity at the putamen level with the tensors matrix for the ber tracking algorithms. Our approach automatically detects the elements that until now were obtained by user interaction: detection of the slice of interest, detection of volumes of interest, automatic detection of the registration parameters. Introducing parameters for ber evaluation and eliminating the demographic factors at the atlas level, as well as at the volume level represents another contribution of PDFibAtl@s.

5.1 Diagnosis evaluation Being able to conrm the cognitive test performed to place the patient on a severity scale is helpful for the medical doctor and oers the possibility to augment the

40

Chapter 5. Conclusion. Further study. Applicability

degree of trust on the diagnosis. Having a test based entirely on the image is a robust and reliable way to evaluate the patient and his current estate. Another important aspect is represented by the fact that the diagnosis is directly liked to the severity of the disease, as it can be cognitively detected and placed only after it passes the second level on the H&Y scale. Having a system based on technical measurable, with a high granularity system oers us the possibility to be apply it on any patient at any level of severity of the disease.

5.2 Prognosis potential The diagnosis and the prognosis are highly dependable as the early diagnosis is not reachable without having the prognosis step dened. This step is reached by evaluating the patients form our database and placing them on an evolutive function. By extending this function and extrapolating towards the low values of the scale, the early estate of the disease is reached. Having the values for this level of the disease oers the information necessary for placing new patients at this level and making diagnosis for these patients as well. Like that we use the prognoses for new diagnosis level.

5.3 PDFibAtl@s importance. Future applicability. Although the prognosis step is not nalized yet, the diagnosis level oers a good correlation between the cognitive scale and the detected values. All the necessary elements for the diagnosis are already determined and extracted automatically. From the technical perspective we have a robust system that encapsulates all the necessary image treatment starting from the scanned images to the motor bers and their density. Automatic detection of the volumes of interest contours an atlas-based method entirely independent on the subject. Even if this approach is specic for the disease, its methods of detection can be applied for other diseases and for any patient. Another very important aspect of our approach is the ber growth algorithm that is based on the anatomical elements: the way the bers are constructed - the WM importance and the way the motor tract is placed inside the human brain. Actually the automatic detection of the volumes is based on the anatomy of the brain and the relative placement of the detected structures in the brain geometry. The limits imposed for the ber tracts are not only for imposing a certain granularity for choosing the bundle of interest, but also to validate the obtained bundle from the anatomical point of view, by the neurologist - validation based on the placement of the bers in the volumes of interest. Rening the ber detection method and making it specic to the gray matter can augment the degree of trust for the diagnosis and add reliability to the system. Also it can oer a higher correlation factor between the score and the detected bers. Diagnosis and prognosis are linked together by the variation function of the bers

5.4. Contribution

41

parameters on the H&Y scale. This function has to be determined on a heterogenous database and should be sensitive only to the bers elements, not to the demographic ones.

5.4 Contribution There are two aspects that reect the contribution of this study: the technical aspect and the medical relevance. The technical aspect represents the algorithms implemented, their originality and their robustness. The medical relevance represents the diagnosis aspect and the prognosis. The volumes of interest are specic for the disease and the manner in which they are detected, by combining the image specic processing methods, together with the geometrical elements and by integrating the anatomy elements. The method is applicable to any patient as it does not take into account the provenance of the case, for shape variability, or the volume of the brain, that variate according to the sex of the subject. This method is for these reasons a complex one, integrating concepts from the medical knowledge for technical purpose. The most important for the originality of the approach is the combination between the automatic volume detection and the ber growth limitation. Also the metrics used for estimating the bers are specic to this approach and are meant not only to evaluate the bers, but also to overcome demographic variation. Using the bers to evaluate the PD evolution is highly reliable, as in previous studies only our rst volume of interest has provided enough data to reach this purpose. Our method adds the 3D aspect in the evaluation by including the bers. The anatomy of the brain incorporates the medical knowledge to the approach, supports the technical elements and is able to link the processing algorithms by oering decisional rules for the detection steps of our system. It is very valuable to be able to take the images from the les directly and by an automatic approach, without considering the identity/demograc information from the patient, to deliver a comprehensive information to the clinician. PDFibAtl@s eliminates the human intervention altogether. At the ROI/VOI and registration level not only the actual registration is automatic, but also the determination of the parameters for this purpose. Having an original take on the DTI image processing and the detection of PD specic volumes adds up to the overall value of the system. The ber algorithm has a new perspective by limiting the bers to the bundle of interest. The lightness of my algorithms is contained in the versatility, as these volume algorithms can be applied on other types of images. For the bers I need the tensors, but these algorithms can be perfected by including a tensor analysis based on the green study presented in chapter4. We can extend this approach to other similar diseases like Alzheimer by determining in the same manner the bundles of interest in this case. Also an automatic atlas of the brain by determining all the anatomical volumes, without any mapping involved so by taking out the demographic importance

42

Chapter 5. Conclusion. Further study. Applicability

can be done based on this approach.

5.5 Conclusion Summarizing the contribution at this level I have developed a system that is able to detect using DTI images of EPI the specic volumes of interest automatically. The algorithms for preprocessing the medical images, removing the skull and the noise represent another important contribution for my system. The fusion that I realized is at the information level and needed registration as well. My contribution at this point is represented by the algorithm that automatically detects the landmarks and coecients for the registration algorithm. At the ber growing algorithm the contribution that I add is contained in using the volumes of interest for the choice of the bundle of interest. Another contribution is represented by the algorithm of statistical analysis and correlation at the validation stage.

List of Abbreviations ADC, 1 Apparent Diusion Coecient, 2

Hohen & Yahr scale of PD severity, i HY, 39

MedINRIA, 37 CSF, 18, 20, 23 tool that makes alignment, ROI deCerebro Spinal Fluid, liquid contection and ber growing, 16 tained inside the brain, 23 MRI, 1 DICOM, 1, 10, 13, 14, 36 Medical Radiography Image, 3 Digital Imaging and Communications in Medicine (DICOM) PD, 2, 39 Parkinson's Disease, i standard for distributing and viewing any kind of medical im- SN, 24, 28, 32 age regardless of their origin, 10 Substantia Nigra - located in the DTI, 1, 3, 9, 10, 12, 15, 39 midbrain area is the producer of Diusion Tensor Imaging, i dopamine, 3 EPI, 5, 11, 14, 17, 18, 21, 24, 28, 29, 31 SOI Slice of Interest - a medical image Echo Planar Imaging, a DTI type of 2D, part of a 3D image stack image that contains tensor inforwhere a region of interest is lomation on all diusion directions cated or part of the VOI, 23 (see2.2), i SPM, 17, 25 Statistical Parameter Mapping, 17 FA, 1, 2, 5, 11, 14, 17, 31, 39 Fractional Anisotropy image obT1 or T2 tained from EPI by computing type of MRI images, 3 the anisotropy value on the diffusion (see 2.3), i UPRS FD, 33 Unied Parkinson Rating Scale Fiber Density, 17 used for rating PD severity, 3 FSL FreeSurfer for Linux integrated in VBM Voxel Based Morphometry system MatLab, 17 that analyzes cerebral images by FV, 33, 34 extending SPM in MathLab, 3 Fiber Volume, 17 VOI, 2224, 29, 31, 36 Volume of Interest, i GM, 18, 23, 24 Grey Matter tissue of the brain WM, 10, 18, 23, 24, 29, 40 sometimes a region, 28 White matter tissue of the brain sometimes a region, 28 H&Y, 3, 29, 33

Appendix A Appendix Dissemination

A.1 Conferences & Workshops 2009 -2010 Teodorescu, R.; Racoceanu, D.; Chan, L.; Lovblad, K. & Muller, H. Parkinson's disease detection using 3D Brain MRI FA map histograms correlated with tract directions - oral presentation Neuroradiology (Brain: Movement and Degenerative Disorders SSC13 - 09) RSNA,95th Radiological Society of North America Scientic Conference and Annual Meeting, November 29 to 4 December, McCormick Place, Chicago IL, USA, 2009. Teodorescu, R.; Racoceanu, D.; Smit, N.; Cretu, V. I.; Tan, E. K. & Chan, L.-L. Parkinson's disease prediction using diusion based atlas - poster session SPIE - Computer Aided Diagnosis [7624-78] PS2, 13-18 Febr., San Diego CA, USA 2010. Teodorescu, R. O. & Racoceanu, D. Prognosis of Parkinson's Disease - poster session, A*STAR Scientic Conference, 28-29 Oct., Biopolis, Singapore 2009. Teodorescu, R. O.; Racoceanu, D. & Chan, L.-L. H&Y compliant for PD detection using EPI and FA analysis - poster session, NIH Workshop Inter-Institute Workshop on Optical Diagnostic and Biophotonic Methods from Bench to Bedside, 1-2 Oct, Washington DC, USA 2009.

A.2 Research stages 2009-2010 February-April 2009 Research stage in Singapore at National University of Singapore (NUS) with the Computer Vision laboratory under the supervision of Prof. Leow Wee Kheng April -October 2009 Research stage in Singapore at Image and Pervasive Access Lab under the supervision of Prof. Daniel Racoceanu from French National Research Center. 18-20 February 2009 Participation at the French-Singaporean symposium at NUS and IPAL Singapore.

Bibliography [Basser 2000] Peter J. Basser, Sinisa Pajevic, Carlo Pierpaoli, Jerey Duda and Akram Aldroubi. In vivo ber Tractography using DT-MRI data. Magnetic Resonance in Medicine, vol. 44, pages 625632, 2000. 19, 29 [Bihan 2001] Denis Le Bihan, Jean-Francois Mangin, Cyril Pupon and Chris A. Clark. Diusion Tensor Imaging : concepts and application. Journal of Magnetic Resonance Imaging, vol. 13, pages 534546, 2001. 19, 29 [Cachier 2003] Pascal Cachier and et al. Iconic feature based non-rigid registration: the PASHA algorithm. Computer Vision and Image understanding, vol. 89, pages 272298, 2003. 26

Multi-contrast large deformation dieomorphic metric mapping for diusion tensor imaging.

[Ceritoglu 2009] Can Ceritoglu, Kenichi Oishi, Xin Li and et al. NeuroImage, vol. 47, pages 618627, 2009. 37

[Chan 2007] L.L. Chan, H. Rumpel and K. Yap. Case-control study of diusion tensor imaging in Parkinsons disease. JNNP, vol. 23, pages 18, 2007. 3, 23, 29

Stanford Study improves insight into Parkinson± disease and possible treatments. Stanford University Medical

[Deisseroth 2009] K. Deisseroth and et al. Centre, vol. 1, page 1, 2009. 18

[Elsevier 2009] Elsevier. Handbook of medical image processing and analysis, second edition (academic press series in biomedical engineering). Academic Press in an imprint of Elsevier, 2009. 26 [Golinpour 2007] Ali Golinpour, Nasser Khetanarvaz and et al. Brain Functional Localization: a Survey of Image Registration Techniques. IEEE Transactions on Medical Imaging, vol. 26, pages 120, 2007. 26, 28 [Kretschmann 2003] Wolfgang Kretschmann Hans-Joachim Weinrich. Cranial neuromiaging and clinical neuroanatomy - 3rd edition. Thieme Classics, 2003. 18

Diusion tensor ber tracking shows distinct corticostriatal circuis in humans. ANN

[Lehericyr 2004] S. Lehericyr, M. Ducros, P.F. Van de Moorele and et al. Neurol, vol. 55, pages 522529, 2004. 3, 18

[Maintz 2000] Antoine JB Maintz and Max A Viergever. A survey of medical image registration. Medical Image Analysis, vol. 1 and 2, pages 132, 2000. 26, 28

48

Bibliography

[Modersitzki 2004] Jan Modersitzki. Numerical methods for image registration (numerical mathematics and scientic computation) (hardcover). Oxford Science Publications, 2004. 16, 37 [news today 2009] Medical news today. Brain bank Appeal aims to double number of brain donors. www.medicalnewstoday.com, March 2009. Parkinsons awareness week 2009, 20-26 April. 2 [PD 2009] PD. Parkinson± disease statistics. on-line, July 2009. http://parkinsonsdisease.emedtv.com/parkinsonease-statistics.html. 2 [Starr 2009] Cindy Starr and George Mandybur. Grant to improve targeting in Parkinsons Surgery. University of Cincinnati neuroscience institute, January 2009. Mayeld Clinic. 24 [Talos 2003] Ion Florin Talos. Slicer 3 Tutorial The SPL PNL Brain Atlas. presentatin, Brigham and Womens Hospital, march 2003. Surgical Planning Laboratory. vii, 25 [Teodorescu 2009a] R. O. Teodorescu, D. Racoceanu and L-L. Chan. HY compliant for PD detection using EPI and FA analysis. Presentation Number: NIH09NIH01-88, 1-2 October 2009. Natcher Auditorium, National Institutes of Health in Bethesda, MD USA. vii, 32, 37 [Teodorescu 2009b] Roxana Teodorescu, Daniel Racoceanu, LL Chan, KO Lovblad and H Muller. Parkinsons disease detection using 3D Brain MRI FA map histograms correlated with tract directions. RSNA, vol. 8015681, page 1, 2009. Chacago, IL USA. ix, 33, 37

Feature extraction and Ontology use for Brain medical images - PhD Report No 1. Rapport technique 1, UPT

[Teodorescu 2009c] Roxana Oana Teodorescu. and UFC, January 2009. 1

[Teodorescu 2009d] Roxana Oana Teodorescu and Daniel Racoceanu. Prognosis of Parkinsons Disease. ASTAR Scientic Conference, 28-29 Oct. 2009. Biopolis, Singapore. vii, 32, 37 [Teodorescu 2010] Roxana Teodorescu, Daniel Racoceanu, Nicolas Smit, Vladimir Ioan Cretu, Eng King Tan and Ling-Ling Chan. Parkinsons disease prediction using diusion based atlas. SPIE Medical Imaging, Computer Aided Diagnosis, San Diego, California, USA, 13-18 Feb. 2010. ix, 34, 35, 37 [Vaillancourt 2009] D. Vaillancourt and et al. Imaging Technology May trace Development of Parkinsons disease. University of Illinois at Chicago, Rush University, vol. 1, page 3, 2009. 3, 18

Bibliography

49

[Vercaunteren 2008a] T. Vercaunteren, X. Penec and et al. Dieomorphic Demons: Ecient Non-parametric Image Registration. NeuroImage, vol. 45, pages 6172, 2008. 16, 37 [Vercaunteren 2008b] Tom Vercaunteren, Xavier Pennec, Aymeric Perchant and Nicholas Ayache. Symmetric Log-Domain Dieomorphic Registration: A Demons - Based Approach. Lecture Notes in Computer Science - Medical Image Computing and Computer-Assisted Intervention -MICCAI, vol. 5241, pages 754761, 2008. 16, 37 [Westin 2002] C.-F. Westin, S.E. Maier and H. Mamata. Processing and visualization for diusion tensor MRI. Medical Image Analysis, vol. 6, pages 93108, 2002. 24 [Wirijadi 2001] Oliver Wirijadi. Survey of 3D Image Segmentation hoer technical report, vol. 1, page 1, 2001. 28

Methods. Fraun-

Cerebral morphology and dopamine D2/D3 receptor distribution in humans: a combined [18F] fallypride and voxel-based morphometry study. NeuroImage, vol. 46, pages

[Woodward 2009] N. Woodward, D. Zald, Z. Ding and et al.

3138, 2009. 3, 18

[Zitova 2003] Barbara Zitova and Jan Flusser. Image Registration: A Survey. Image and Vision Computing, vol. 21, pages 9771000, 2003. 17, 28

Design and Use of Anatomical Atlases for Radiotherapy

Image. Analysis,. Medical Image. Processing,. PD. Detection,. Prediction ...... We have tested our images on 3DSlicer, a multi-platform free open source soft-.

3MB Sizes 1 Downloads 275 Views

Recommend Documents

Design and Use of Anatomical Atlases for Radiotherapy
codes (MLC) and bit-interleaved coded modulation (BICM) where the ...... ming code, Golay code, convolutional code, Bose-Chaudhuri-Hocquenghem (BCH) ...

Combination of Trastuzumab and Radiotherapy ... - OMICS International
Feb 28, 2018 - Overall Survival (OS) [1]. Many clinical trials and meta-analyzes have standardized the use of anthracyclines and Taxanes in the adjuvant.

OPTIMIZATION OF INTENSITY-MODULATED RADIOTHERAPY ...
NTCPs based on EUD formalism with corresponding ob- Fig. 1. (a) Sample DVH used for EUD calculation. (b) EUD for the. DVH in (a) as a function of parameter a. Tumors generally have. large negative values of a, whereas critical element normal struc- t

OPTIMIZATION OF INTENSITY-MODULATED RADIOTHERAPY ...
deviates from EUD0. For a tumor, the subscore attains a. small value when the equivalent uniform dose falls sig- nificantly below EUD0. Similarly, for a normal ...

Deformable Atlases for the Segmentation of Internal ...
ical knowledge in the shape of digital atlases that deform to fit the image data to be analysed. ... In other MRI sequences, like the Fluid Attenuated Inversion Recovery (FLAIR) sequence, the CSF is .... It uses a block matching strategy in a two-.

Radiotherapy III.pdf
Connect more apps... Try one of the apps below to open or edit this item. Radiotherapy III.pdf. Radiotherapy III.pdf. Open. Extract. Open with. Sign In. Main menu.

Effects of radiotherapy with concomitant and adjuvant ...
Mar 9, 2009 - and 9·8% (6·4–14·0) at 5 years with temozolomide, versus 10·9% (7·6–14·8), 4·4% .... radiotherapy or standard radiotherapy plus concomitant.

Effects of radiotherapy with concomitant and adjuvant ...
Mar 9, 2009 - gave modest further benefit: a meta-analysis of ... putative prognostic and predictive factors. .... analyses on few patients lack statistical power.

Medical Radiotherapy III.pdf
Malignant pleural effusion management. **********************. Page 1 of 1. Medical Radiotherapy III.pdf. Medical Radiotherapy III.pdf. Open. Extract. Open with.

Radiotherapy II.pdf
Page 1 of 1. Radiotherapy II.pdf. Radiotherapy II.pdf. Open. Extract. Open with. Sign In. Main menu. Displaying Radiotherapy II.pdf.

Radiotherapy I.pdf
7. Integral dose and its implications. 8. Ir 192. 9. Photo dynamic therapy. ****************************. Page 1 of 1. Radiotherapy I.pdf. Radiotherapy I.pdf. Open.

A stand-alone method for anatomical localization of ...
implemented as part of our NIRS Analysis Package (NAP), a public domain Matlab ... MNI space (Okamoto et al., 2004; Okamoto and Dan, 2005; Singh et al.,.

Anatomical Coupling between Distinct Metacognitive Systems for ...
Jan 30, 2013 - functionally distinct metacognitive systems in the human brain. Introduction. What is the ..... define regions of interest (ROIs) using MarsBar version 0.42 software .... model for finding the best-fit line while accounting for errors

Medical Radiotherapy II.pdf
Short essays: (8x10=80). 2. PAP smear. 3. Para neoplastic disorders. 4. Glioblastoma multiforme. 5. Wilm's tumor. 6. Oncogenes. 7. Tumor markers. 8. Apoptosis.

Medical Radiotherapy I.pdf
I 131. 6. Ophthalmic brachytherapy applications in clinical practice. 7. Photo dynamic therapy. 8. Biological effects seen after an accidental whole body exposure ...

Medical Radiotherapy II.pdf
Describe the epidemiology and aetiology of breast cancer. 7. Pathological features of basal cell carcinoma and its management. 8. Genetic syndromes in cancer.

Medical Radiotherapy III.pdf
... therapy in prostate cancer. 8. Choriocarcinoma. 9. Tomotherapy. **********************. Page 1 of 1. Medical Radiotherapy III.pdf. Medical Radiotherapy III.pdf.

Medical Radiotherapy III.pdf
Discuss the management of Ewing's sarcoma. Short essays: (8x10=80). 2. Management of oncological emergencies. 3. Tyrosine kinase inhibitors. 4. P32. 5.

Automated Segmentation and Anatomical Labeling of ...
Automated Segmentation and Anatomical. Labeling of Abdominal Arteries Based on Multi-organ Segmentation from Contrast-Enhanced CT Data. Yuki Suzuki1 ...

The use of EEG-based inverse models for both BCI design and brain ...
Open-ViBE : An open-source software for Brain-Computer. Interfaces (BCI) and Virtual Reality (VR) ... BCI design using an inverse model. ▫ Conclusion ... Data set IV of the « BCI competition 2003 » [Blankertz et al, 04]. ▫ Protocol : self paced

Medication - Use of Unlicensed Medicines and Off-label Use of ...
There was a problem previewing this document. Retrying... Download. Connect more apps... Try one of the apps below to open or edit this item. Medication - Use of Unlicensed Medicines and Off-label Use of Licensed Medicines.pdf. Medication - Use of Un