Diagnosing Adenocarcinoma of the Prostate by Computer Vision Methods
Ravishankar Sivalingam, Guruprasad Somasundaram, Xinyan Li, Jonathan C Henriksen, Arindam Banerjee, Vassilios Morellas, Nikolaos Papanikolopoulos, Alexander M Truskinovsky. University of Minnesota, Minneapolis, MN
Background: Based on our successful application of computer vision methods to detect endometrial carcinoma, we extend our analysis to acinar-type adenocarcinoma of the prostate, using region covariance descriptors in a dictionary-learning modeling framework and discriminative dictionary learning.
Design: We used 60 images of H&E-stained sections of radical prostatectomy specimens from 3 patients, scanned at x50 magnification on a digital slide scanner. The color images were transformed to grayscale using a custom transformation and then manually annotated to train the classification algorithm. The annotated regions are then broken down into overlapping blocks of 200 x 200 pixels. Each block is represented by the covariance matrix over the image features in that block. We used a set of spatial (x,y,ρ,θ) and intensity [I, Ix, Iy, √(Ix2 + Iy2)] features, giving rise to 8x8 covariance descriptors. The covariance descriptors from each class (cancer vs. benign) were used to learn a concise dictionary model. The experiments were repeated for varying dictionary size K and sparsity level T. Each test block was classified as being cancerous or benign based on which dictionary model gave the least representation error. 10-fold cross-validation was run on the entire dataset.
In discriminative dictionary learning, 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. 32 x 32 patches were extracted from images of each class and discriminative dictionaries were learned using 12000 labeled training patches, followed by three-fold cross-validation.
Results: Using region covariance descriptors, the overall classification accuracy of the image blocks into cancerous vs. benign regions was 88.24%. On average, we obtained a true positive rate of 98.3% and a false positive rate of 21.8%, with a standard deviation of < 1%. These results were obtained for dictionary size of K = 1600 and sparsity of T = 4, which were selected by cross-validation. In contrast, discriminative dictionary learning, which had performed well in detecting endometrial carcinoma, did not produce sufficient discrimination between the benign and cancerous regions in the prostate images, with classification accuracy close to a random classifier (∼50%).
Conclusions: Computer vision is highly effective in diagnosing adenocarcinoma of the prostate. As different types of cancer have very different image characteristics, custom classification approaches should be developed for individual tumors.
Tuesday, March 20, 2012 1:00 PM
Poster Session IV # 220, Tuesday Afternoon