Robust Detection of Renal Calculi from Non-contract CT Images Using TV-flow and MSER Features

a

Jianfei Liua, Shijun Wanga, Marius George Lingurarub, Ronald M. Summersa Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, National Institutes of Health Clinical Center, Bethesda, MD USA 20892-1182; b Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Medical Center, Washington, DC, USA 20010 ABSTRACT

Renal calculi are one of the most painful urologic disorders causing 3 million treatments per year in the United States . The objective of this paper is the automated detection of renal calculi from CT colonography (CTC) images on which they are one of the major extracolonic findings. However, the primary purpose of the CTC protocols is not for th e detection of renal calculi, but for screening of colon cancer. The kidneys are imaged with significant amounts of noise in the non-contrast CTC images, which makes the detection of renal calculi extremely challenging. We propose a computer-aided diagnosis method to detect renal calculi in CTC images. It is built on three novel techniques: 1) tota l variation (TV) flow to reduce image noise while keeping calculi, 2) maximally stable extremal region (MSER) features to find calculus candidates, 3) salient feature descriptors based on intensity properties to train a support vector machin e classifier and filter false positives. We selected 23 CTC cases with 36 renal calculi to analyze the detection algorithm . The calculus size ranged from 1.0mm to 6.8mm. Fifteen cases were selected as the training dataset, and the remaining eight cases were used for the testing dataset. The area under the receiver operating characteristic curve (AUC) values were 0.92 in the training datasets and 0.93 in the testing datasets. The testing dataset confidence interval for AUC reported by ROCKIT was [0.8799, 0.9591] and the training dataset was [0.8974, 0.9642]. These encouraging result s demonstrated that our detection algorithm can robustly and accurately identify renal calculi from CTC images. Keywords: Renal calculi detection, Total variation flow, Maximal stable extremal region, CT colonography

1. INTRODUCTION Renal calculi are one of the most painful urologic disorders and approximately 12% of men and 5% of women in th e United States will be affected during their life. Calculi can occur at any age and are more likely to recur within five year s after the first occurrence. Each year nearly 3.3 million Americans require medical care for renal calculus removal and pain relief at a cost of $5.3 billion per year. Unsurprising, renal calculi are regarded as one of the important extracoloni c findings at CT colonography (CTC)1. Therefore, early detection of renal calculi on CTC is clinically useful to decreas e the chance of causing severe pain as well as reduce the treatment cost. The primary purpose of the CTC protocols i s unfortunately not for the detection of renal calculi, but for screening for colon cancer. The kidneys are imaged with significant amounts of noise in the non-contrast CTC images, which makes it difficult to accurately detect renal calculi a t CTC images. Despite the fact that early detection of renal calculi is clinically useful, there are few works on this important topic . Tamilselvi2 developed a semi-automatic region growing algorithm on ultrasound images to find calculi candidates and established texture features on the segmented candidates. The existence of the calculi is detected by choosing the spatia l gray level dependence method to evaluate the texture features. The input image was finally classified into thre e categories, normal, early detection (small stones that do not cause symptoms), and kidney stones. However, this process involves extensive manual operations. Later on, Tamilselvi3,4 improved their approach with a learning process called adaptive neuro fuzzy inference system (ANFIS) to classify the selected calculi with texture features as well as a contou r based method to enhance the segmentation accuracy. A similar approach to detect calculi on ultrasound images can b e 6 found in Shah’s work except that more comprehensive texture features are constructed for classification. Lee et al.5 proposed a computer-aided diagnosis system to differentiate urinary stone and vascular calcifications on pre-contrast CT images, which is similar to our work. In their method, they semi-automatically chose calculi candidates using region growing followed by statistical and shape feature computation. Urinary stone and vascular calcifications are finally d istinguished by an artificial neural network. Nevertheless, existing work are mainly concerned with the detection of renal calculi on ultrasound images, and they all required large amounts of manual operations. Medical Imaging 2013: Computer-Aided Diagnosis, edited by Carol L. Novak, Stephen Aylward, Proc. of SPIE Vol. 8670, 867006 · © 2013 SPIE · CCC code: 1605-7422/13/$18 · doi: 10.1117/12.2008034

Proc. of SPIE Vol. 8670 867006-1

In this paper, we propose, to our knowledge, the first fully automatic renal calculi detection algorithm on non-contrast CT images. Our approach has three novel contributions: 1) total variation (TV) flow8 to reduce image noise while preserving calculi, 2) maximally stable extremal region (MSER) feature9 to find calculus candidates, 3) salient feature descriptors based on intensity properties to train a support vector machine classifier and filter false positives. We validate our algorithm on 23 CTC datasets split into training and testing datasets, and experimental results revealed that our detection algorithm could accurately detect renal calculi larger than 1mm.

2. METHODOLOGY Our detection approach contains four main steps, including kidney segmentation, TV-flow smoothing, calculi candidate detection and segmentation, and false positive reduction. In this section, we describe our algorithm in detail. 2.1 Kidney Segmentation This step extracts kidneys to limit the search ranges of renal calculi because stones are located inside the kidney. Left and right kidneys are automatically segmented from CTC images by using a shape-prior constrained segmentation method7. Liver and spleen are first segmented from the current patient image10 and they are used to define the possible kidney spatial ranges. Five reference CT images and their manual segmented kidneys are registered with the current patient image to build a probabilistic atlas. The atlas is then imported into an energy function represented as a Markov random field, and efficient belief propagation11 is chosen to minimize the energy function, which yields the final kidney segmentation, as illustrated in Fig. 1.

(a) original CT

(b) organ segmentation

(c) 3D visualization

Figure 1. Illustration of kidney segmentation, where green objects represent the segmented liver, blue objects denote spleen, and light brown objects correspond to kidneys. 2.2 TV-Flow Smoothing The sub-images containing segmented kidneys can be extracted from the original CT image, as illustrated in Fig. 2(b). Thus, the search space of the renal calculi is significantly reduced. However, kidney regions are usually imaged with large amounts of noise in the CTC images. We choose total variation (TV) flow8 to reduce image noise, which is a nonlinear diffusion process attempting to denoise an initial image I with the partial differential equation

∂ t u = div(g ( ∇u )∇u ) g ( ∇u ) =

1 ∇u

(1)

u (t = 0) = I The fundamental idea of TV-flow is to minimize the global total variation

∫

Ω

∇u dx of the result in the image domain

Ω while not deviating too much from the original signal. The diffusion process is controlled by the diffusivity function,

Proc. of SPIE Vol. 8670 867006-2

g ( ∇u ) . The boundaries between calculi and tissues are thus well preserved as the calculi boundaries correspond to the image edges and the diffusion process stops at these image edges. The diffusivity function described by L1 norm in the TV flow is also demonstrated to efficiently handle both Gaussian and non-Gaussian noise8. No additional parameters are required to be optimized in

g ( ∇u ) . More important, experiments reveal that the diffused results are close to piecewise

constant segmentation, as shown in Fig. 2(c). Image noise in the non-calcified regions is substantially reduced, while the calculus is well kept. This desirable property makes it tractable to segment renal calculi.

.-.

Figure 2. Process of MSER feature detection on CTC images. (a) Original CT image, (b) Sub-image containing right kidney, (c) Smoothed sub-image using TV-flow, (d) Detected MSER demonstrating the calculus. During numerical computation, Eq. 1 is transformed into

u k +1 − u k

τ

= A(u k )u k +1

where k is the iteration index, A(uk) is a diffusion matrix defined by the diffusivity function

 g i ↔ j if j ∈ N (i )  if j = i aij = − ∑n∈N (i ) g i ↔ j  0 else 

Proc. of SPIE Vol. 8670 867006-3

(2)

g ( ∇u )

,

(3)

Here,

g i ↔ j represents the diffusivity between the pixels i and j, and N (i ) denotes the neighboring pixels of i. Additive

operator splitting (AOS) scheme12 is selected to compute Eq. 2 for its efficiency on handling diffusion computation. The essential idea is to subdivide the diffusion process along each image dimension. Letting the image dimension be D and the current image direction be l-direction, Eq. 2 is derived as

(1 − DτA l (u k ))ul where 1 is the identity matrix, and

ul

k +1

k +1

= uk

is the diffused value along l-direction.

(4)

A l (u k ) measures the diffusion

between neighboring pixels only along l-direction. AOS-scheme uses the Thomas algorithm13 to resolve Eq. 4, and the final solution is represented as

u k +1 =

1 D k +1 ∑ ul D l =1

(5)

2.3 Calculi Candidate Detection and Segmentation The characteristic of renal calculi is their high intensity values in comparing with surrounding tissues as shown in Fig. 2(c). Maximally stable extremal region (MSER)9 feature is well suited for the detection of renal calculi since it exploits the extreme intensity values to detect blob features.

Q ⊂ R N is called an extremal region if for all points p ⊂ Q and all boundary points q ⊂ ∂Q , I ( p ) > I (q ) or I ( p ) < I (q ) . Let Q1 ,  , Qi −1 , Qi ,  be a sequence of Assuming Ω =RN be an N-dimensional image domain, a region

nested extremal regions

Qi ⊂ Qi +1 . Extremal region Qi is maximally stable if and only if f (i ) = Qi +∆ \ Qi −∆ / Qi

has a local minimum. The process of MSER feature detection is thus essentially a thresholding process, and an image region is considered as a MSER feature candidate if it can exist within a threshold range i − ∆, i + ∆ .

[

]

9

However, the original MSER feature detection is developed for 2D images, and we extend the MSER algorithm to 3D CTC images by finding maximally stable extremal volume. We select feature points with HU value larger than 100 that belong to renal calculi. Moreover, geodesic active contours14 were exploited to replace region growing in the original MSER algorithm to extract renal calculi due to their superior accuracy to segment data of high texture. Finally, we perform the modified MSER algorithm on the smoothed sub-image, and the detected MSER is illustrated in Fig. 2(d), which corresponds exactly to the renal calculus in Fig. 2(c). 2.4 False Positive Reduction SVM classifier15 is used to filter false positives from the MSER feature detection. We build a four-dimensional feature vector to train the SVM classifier, including the mean intensity value of the calculi candidate, the standard deviation, the intensity difference between the candidate and its surrounding object, and the distance of the current MSER to the kidney boundary. The renal calculi are finally detected based on the decision of the SVM classifier.

3. EXPERIMENTAL RESULTS We selected twenty-tree CTC cases with 36 renal calculi total to validate our algorithm. 20 calculi are located in the right kidney and the remaining calculi in the left kidney. All calculi are annotated by an experienced radiologist. Retrospective analyses of these images were approved by our Office of Human Subjects Research. The slice thickness is 1mm. The calculus size ranged from 1.0mm to 6.8mm. Fifteen cases with 27 calculi were selected to build the training datasets, and the remaining cases were used for the testing datasets. Fig. 3 gives the receiver operating characteristic (ROC) curves on the training and testing dataset. The area under curve (AUC) was 0.92 in the training dataset and 0.93 in the testing dataset. The 95% confidence interval reported by ROCKIT16 was [0.8974, 0.9642] in the training dataset and [0.8799, 0.9591] in the testing dataset. These promising results demonstrated that our detection algorithm can accurately detect renal calculi from CTC images.

Proc. of SPIE Vol. 8670 867006-4

AUC = 0.9235

AUC = 0.9306

0.9

0.9

0.8

0.8

0.7

7+

a--

Z2 0.6

ÿ 0.5

0.7

'5

0.6

:.= (73

0.

0.4

0.4

0)

0.3

0.3

0.2

0.2

0.1

0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False Positive Rate

False Positive Rate (a) Training datasets

(b) Testing datasets

Figure 3. ROC curves on training and testing datasets. Note that AUC values are above 0.9 on both datasets.

(a) True Positive

(b) False Positive

"

(c) True Negative (d) False Negative Figure 4: Typical examples of true positive, false positive, true negative, and false negative from four patients, where the detected renal calculus candidates are marked with a red square. The response values of SVM classifier are -17.5, -5.2, 5.2, and 1.7, respectively. The smaller the response value, the higher confidence the candidates are calculus. Fig. 4 presents typical examples of true positive, false positive, true negative, and false negative from our detection algorithm. In Fig. 4(a), a 6.8mm calculus marked by a red rectangle was accurately identified with strong SVM response (-17.5) as its intensity values are significantly larger than its surrounding tissue. Fig. 4(b) illustrates a false positive caused by CT image artifacts from contrast material in bowel adjacent to the kidney. In Fig. 4(c), the image artifacts also i ncreased the intensity values inside the marked square. However, the increase was small in comparison with Fig. 4(b), and our detector was able to exclude it. Fig. 4(d) shows a challenging example. The intensity values of the calculus are nearly indistinguishable from the adjacent tissue, which causes our detection algorithm to miss it.

Proc. of SPIE Vol. 8670 867006-5

4. CONCLUSION AND FUTURE WORK In this work, we describe a new fully-automated method to detect renal calculi, one common type of extracolonic finding, from non-contrast CTC images. To our knowledge, the detection of renal calculi on the challenging CTC images was not studied in the field of computer-aided diagnosis before since CTC images are used primarily for the detection of colon cancer. Total variation flow, maximally stable extremal region, and salient feature descriptor for SVM classifier are three key contributions to address this issue. The proposed method tested on 23 cases showed that the AUC values were 0.92 in the training datasets and 0.93 in the testing datasets. The nearly identical AUCs on training and testing suggest that the algorithm will be generalizable to fresh datasets and that overtraining did not occur. These encouraging results demonstrated that our detection algorithm is able to robustly and accurately detect renal calculi from CTC images. In the future, we will investigate more advanced texture features and shape features as used in the existing approaches3,4,5,6. In addition, we are collecting more CTC datasets to validate our renal calculi detection algorithm.

ACKNOWLEDGEMENTS This work was supported by the Intramural Research Program of the National Institutes of Health, Clinical Center.

REFERENCES [1] Pickhardt, P.J., Hanson, M.E., Vanness, D.J., Lo, J.Y., Kim, D.H., Taylor, A.J., Winter, T.C., and Hinshaw, J.L., " Unsuspected Extracolonic Findings at Screening CT Colonography: Clinical and Economic Impact," Radiology, 249(1), 151-159 (2008). [2] Tamilselvi, Ms.P.R. and Thangaraj, Dr.P., “Computer Aided Diagnosis System for Stone Detection and Early Detection of Kidney Stones,” Journal of Computer Science, 7(2), 250-254 (2011). [3] Tamilselvi, P.R. and Thangaraj, P., “An Efficient Segmentation of Calculi from US Renal Calculi Images Using ANFIS System,” European Journal of Scientific Research, 55(2), 323-333 (2011). [4] Tamilselvi, P.R. and Thangaraj, P., “Segmentation of Calculi from Ultrasound Kidney Images by Region Indicator with Contour Segmentation Method,” Global Journal of Computer Science and Technology, 11(22), (2011). [5] Lee, H.J., Kim, K.G., Hwang, S.II, Kim, S.H. Byun, S.S., Lee, S.E., Hong, S.K., Cho, J.Y., and Seong, C.G., “Differentiation of Urinary Stone and Vascular Calcifications on Non-contrast CT Images: An Initial Experience using Computer Aidied Diagnosis,” Journal of Digital Imaging, 23(3), 268-276 (2010). [6] Shah, S.R., Desai, M.D., and Panchal, L., “Identification of Content Descriptive Parameters for Classification of Renal Calculi,” International Journal of Signal and Image Processing, 1(4), 255-259 (2010). [7] Liu, J., Linguraru, M.G., Wang, S., and Summers, R.M., “Automatic Segmentation of Kidneys from Non-contrast CT Images Using Efficient Propagation,” Accepted by SPIE Medical Imaging 2013. [8] Brox, T. and Weickert, J., “A TV flow based local scale estimate and its application to texture discrimination,” Journal of Visual Communcation and Image Representation, 7(5), 1053-1073 (2006). [9] Matas, J., Chum, O., Urban, M., and Pajdla, T., "Robust wide baseline stereo from maximally stable extremal regions," Proc. British Machine Vision Conference, 384-396 (2002). [10] Linguraru, M.G., Sandberg, J.K., Li, Z., Shah, F., and Summers, R.M., “Automated segmentation and quantification of liver and spleen from CT images using normalized probabilistic atlases and enhancement estimation,” Medical Physics, 37(2), 771-783 (2010). [11] Felzenszwalb, P.F. and Huttenlocher, D.P., “Efficient Belief Propagation for Early Vision,” International Journal of Computer Vision, 70(1), 41-54 (2006). [12] Weickert, J., ter Haar Romeny, B.M., and Viergever, M.A., “Efficient and reliable schemes for nonlinear diffusion filtering,” IEEE Transactions on Image Processing, 7(3), 398-410 (1998).

Proc. of SPIE Vol. 8670 867006-6

[13] Press, W.H., Teukolsky, S.A., Vetterling, W.T., and Flannery, B.P., “Numerical Recipes 3rd Edition: The Art of Scientific Computing,” Cambridge University Press (2007). [14] Caselles, V., Kimmel, R., and Sapiro, G., “Geodesic Active Contours,” International Journal of Computer Vision, 22(1), 61-97 (1997). [15] Cortes, C. and Vapnik, V., “Support-Vector Networks,” Machine Learning, 20, 273-297 (1995). [16] Metz, C.E., Herman, B.A., and Shen, J.H., “Maximum likelihood estimation of receiver operating characteristic (ROC) curves from continuously-distributed data,” Statistics in Medicine, 17, 1033-1053 (1998).

Proc. of SPIE Vol. 8670 867006-7