22nd International Conference on Pattern Recognition (ICPR), Stockholm, Sweden, 24-28 August, 2014
Automatic CAD System for HEp-2 Cell Image Classification Shahab Ensafi1,2, Shijian Lu1, Ashraf A. Kassim2, Chew Lim Tan3 1Institute
for Infocomm Research (I2R), Agency for Science, Technology and Research (A*Star), Singapore 2Department of Electrical and Computer Engineering (ECE), National University of Singapore (NUS), Singapore 3School of Computing (SoC), National University of Singapore (NUS), Singapore Email:
[email protected],
[email protected]
Abstract
Dataset
It has been estimated that autoimmune diseases are among the top ten leading causes of death among women in all age groups up to 65 years. However, the detection of it by indirect immunofluorescence (IIF) image analysis depends heavily on the experience of the physicians. An accurate and automatic Computer Aided Diagnosis (CAD) system will help greatly for the classification of the Human Epithelial type 2 (HEp-2) cell images with little human intervention. In this work we present an automatic HEp-2 cell image classification technique that exploits different spatial scaled image representation and sparse coding of SIFT features. Additionally, spatial max pooling of sparse coding at different scales is used to boost the classification performance. The proposed method is tested on the ICPR 2012 contest dataset and experiments show that it clearly outperforms state-of-the-art techniques in cell and image level as well as two intensity level images.
The publicly available MIVIA HEp-2 dataset (ICPR2012) [4] is used.
Centromere (Ce)
Coarse speckled (Cs)
Cytoplasmatic (Cy)
Fine speckled (Fs)
Homogeneous (H)
Nucleolar (N)
ICPR2012 Dataset Centromere Coarse speckled Cytoplasmatic Fine speckled Homogeneous Nucleolar
Introduction Diagnosing the Autoimmune Diseases (Ad) based on Indirect Immunofluorescence (IIF) imaging technique is widely used. With an increasing number of the AD occurrence, an automated Computer Aided Diagnosis (CAD) system is required, for the benefits of:
Ce
Training set
1. Lower cost
Cy
Test set
Fs
H
Total
Intermediate
Positive
Intermediate
Positive
Intermediate
Positive
2 (119) 1 (41) 1 (24) 1 (48) 1 (47) 1 (46)
1 (89) 1 (68) 1 (34) 1 (46) 2 (103) 1 (56)
1 (65) 1 (33) 1 (13) 1 (63) 1 (61) 1 (66)
2 (84) 2 (68) 1 (38) 1 (51) 1 (119) 1 (73)
3 (184) 2 (74) 2 (37) 2 (111) 2 (108) 2 (112)
3 (173) 3 (136) 2 (72) 2 (97) 3 (222) 2 (129)
7 (325) 7 (396) 14 (721)
Total
Cs
6 (301) 8 (433) 14 (734)
13 (626) 15 (829) 28 (1455)
N Overall 6 (257) 5 (210) 4 (109) 4 (208) 5 (330) 4 (241) 28 (1455)
Results and discussion
2. Faster diagnosis 3. Repeatability of diagnosis
Cell and Image Level Classification
Some of the previous methods: Co-occurrence among Local Binary Pattern (CoALBP) based features with Support Vector Machine (SVM) classifier [1]. Frequency histogram of textons on top of the texture features and a k-NN with 𝜒 2 distance classifier [2]. However, the performance of the classification of cell patterns is not as good as human experts.
Cell Level: all the cells separately, without their surrounding cells, are classified. Image level: all cells in one specimen image is classified with assumption that all the cells in one image belong to one class.
100 90 80 70 60 50 40 30 20 10 0
cell level Image level
Method Proposed Method
Ce
Cs
Cy
Fs
H
N
Ce
88.6
0.0
0.0
0.0
0.0
11.4
Cs
6.9
62.4
4.0
20.8
5.0
Cy
0.0
2.0
98.0
0.0
Fs
4.4
28.9
0.9
H
1.7
1.1
N
6.5
4.3
Ce
Cs
Cy
Fs
H
N
Ce
100.0
0.0
0.0
0.0
0.0
0.0
1.0
Cs
0.0
66.7
0.0
33.3
0.0
0.0
0.0
0.0
Cy
0.0
0.0
100.0
0.0
0.0
0.0
29.8
36.0
0.0
Fs
0.0
50.0
0.0
50.0
0.0
0.0
0.0
15.0
82.2
0.0
H
0.0
0.0
0.0
0.0
100.0
0.0
1.4
0.7
11.5
75.5
N
0.0
0.0
0.0
0.0
0.0
100.0
Cell Level Confusion Matrix
Image Level Confusion Matrix
Cell Classification Using Intensity Levels
1. The proposed method consists of two stages. In the training stage, the dictionary is learned by using the Grid SIFT features which are samples from all the training images. 2. The sparse coding method is then applied to learn the dictionary and sparse codes iteratively. In the next stage to generate the feature vectors, max pooling is performed on the histogram of the local descriptors in three different scales. 3. Finally a multiclass Support Vector Machine (SVM) is learned for image classification. In the testing stage, the same protocol is performed. The sparse codes are obtained by using the learned dictionary and the classification is done by using the trained SVM model.
90 80 70 60 50 40 30 20 10 0
A grid mesh of equal spacing and a patch around each point. The histogram of SIFT descriptors in each patch is calculated.
All cells Intermediate intensity Positive intensity
Discussion
Features
By gaining the prior knowledge on intensity level of the cell images, two dictionaries are learned and the classification accuracies are achieved.
On the cell level an accuracy of 72.8% is obtained, which is very close to the human expert accuracy at 73.3%. On the image level, the accuracy of 85.8%, which is almost the same as the accuracy that Nosaka’s [21] method and human expert is obtained. The accuracy in the intermediate intensity level is 62.46%, which is almost 13% more than the human expert accuracy (49.5%) and almost 5% more than the best results reported in the ICPR contest. For the positive intensity images, the human expert accuracy is 79.5% and we achieved 81.5%, which improves the results for 2%. Overall, the cell level accuracy by considering the intensity level of the cell images is achieved 73.57%, which is better than the human expert result.
Descriptor Representation
Conclusion
The Bag of Words (BoW) model is used to calculate the sparse codes of the input features. The Idea is to learn an over complete dictionary with 1024 dimension, which can reconstruct the input image with the use of sparse number of dictionary words [3]. Let 𝔽 be a set of features in a 𝒟-dimensional space: 𝔽 = 𝐹1 , 𝐹2 , ⋯ , 𝐹𝑁
𝑇
𝑁
min 𝑍,𝔻
𝐹𝑛 − 𝑧𝑛 𝔻
2
+ 𝜆 𝑧𝑛
Sparse Codes: Z = z1 , z2 , ⋯ , zN
𝑛=1
𝐷𝑘 ≤ 1, ∀𝑘 = 1, 2, ⋯ , 𝐾
s.t.
∈ ℝ𝑁×𝒟
Dictionary:
𝑇
𝔻 = 𝐷1 , 𝐷2 , ⋯ , 𝐷𝐾
𝑇
In this work an automatic CAD system is proposed. Because of the sparse nature of the patches in Grid SIFT calculation and low reconstruction error of sparse coding scheme, this method obtains superior performance in comparison with vector quantization method. After sparse coding, the max pooling of scaled images in three scales is used. For classification, a multiclass linear SVM is used, which is trained by one-versus-all strategy on training images. The proposed method is experimented on publicly available MIVIA HEp-2 images dataset in both cell and image level. Moreover, the classification using two intensity levels is experimented in this work.
References
Classifier Multiclass linear SVM is used: Six linear SVM are learned and combined together to minimize the weights by means of modified Hinge loss (𝓁). 𝑛
𝑦 = max < 𝒘𝑐 , 𝒛 > 𝑐∈𝒴
min 𝐽 𝒘𝑐 = 𝒘𝑐 𝒘𝑐
2
𝓁(𝒘𝑐 ; 𝑦𝑐𝑖 , 𝒛𝑖 )
+𝐶 𝑖=1
[1] R. Nosaka and K. Fukui, “Hep-2 cell classification using rotation invariant co-occurrence among local binary patterns,” Pattern Recognition, 2013. [2] X. Kong, K. Li, J. Cao, Q. Yang, and L. Wenyin, “Hep-2 cell pattern classification with discriminative dictionary learning,” Pattern Recognition, 2013. [3] J. Yang, K. Yu, Y. Gong, and T. Huang, “Linear spatial pyramid matching using sparse coding for image classification,” in Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009, pp. 1794–1801. [4] P. Foggia, G. Percannella, P. Soda, and M. Vento, “Benchmarking hep-2 cells classification methods,” Medical Imaging, IEEE Transactions on, vol. 32, no. 10, pp. 1878–1889, 2013.
22nd International Conference on Pattern Recognition (ICPR), Stockholm, Sweden, 24-28 August, 2014