Diagnosing Endometrial Carcinoma via Computer-Assisted Image Analysis.
Ravishankar Sivalingam, Guruprasad Somasundaram, Aravind Ragipindi, Arindam Banerjee, Alexander M Truskinovsky. University of Minnesota, Minneapolis
Background: The diagnosis of endometrial endometrioid carcinoma is based primarily on architecture. We investigated whether endometrial carcinoma can be distinguished from proliferative endometrium using 2 methods of computer-assisted analysis, 1) region covariance descriptors and 2) discriminative dictionaries.
Design: We used 4 images each of H&E-stained sections of 2 diagnostic classes, proliferative endometrium and well-differentiated endometrioid carcinoma of the endometrium, scanned at x50 magnification on a digital slide scanner.
In the first method, an image region is represented using the covariance matrix over features in the region. Test image patches are classifi ed via k-NN classi fication using a geodesic distance measure over covariances. From the images in each class, we randomly sampled 100 k x k blocks from manually demarcated diagnostic regions in each image. We used 2 types of covariances:
Cwave: 40 x 40 covariances computed over the responses of 10 Daubechies wavelet filters;
Csi: 8 x 8 covariance descriptors computed over spatial (x, y, ρ, θ) and intensity I, Ix, Iy, √([Ix]2 + [Iy]2) features.
We ran 2 types of cross-validation, a) 10-fold cross-validation on the entire dataset and b) image-wise cross-validation to test performance on unseen images, where we randomly select one test image from each class, using the rest for training.
In the second method, we vectorize √n x √n image patches into n-vectors, and a representative dictionary is learned for vectors from each class, under the constraint that each vector can be reconstructed by a sparse subset of the dictionary. The discriminative power lies in the fact that the dictionary from one class performs poorly at reconstructing patches from the other class. 32 x 32 patches were extracted from images of each class and discriminative dictionaries were learned using 12000 labeled training patches. Then three-fold cross-validation was performed.
Results: For the first method, the accuracy of the classification of images into proliferative endometrium and carcinoma was 96-98% for 10-fold cross-validation and 94-96% for image-wise cross-validation. The accuracy was highest for 200 x 200 blocks, showing that this is a discernible scale for classi fication. Csi performed slightly better than Cwave, though the di fference was not statistically signi ficant. For the second method, the overall classi fication accuracy was 60-80%.
Conclusions: Computer-assisted analysis can be used with high accuracy to distinguish well-differentiated endometrioid carcinoma from proliferative endometrium.
Tuesday, March 1, 2011 1:00 PM
Poster Session IV # 193, Tuesday Afternoon