Gland Segmentation and Computerized Gleason Grading of Prostate Histology by Integrating Low-, High-level and Domain Specific Information

In this paper we present a method of automatically detecting and segmenting glands in digitized images of prostate histology and to use features derived from gland morphology to distinguish between intermediate Gleason grades. Gleason grading is a method of describing prostate cancer malignancy on a numerical scale from grade 1 (early stage cancer) through grade 5 (highly infiltrative cancer). Studies have shown that gland morphology plays a significant role in discriminating Gleason grades. We present a method of automated detection and segmentation of prostate gland regions. A Bayesian classifier is used to detect candidate gland regions by utilizing low-level image features to find the lumen, epithelial cell cytoplasm, and epithelial nuclei of the tissue. False positive regions identified as glands are eliminated via use of domain-specific knowledge constraints. Following candidate region detection via low-level and empirical domain information, the lumen area is used to initialize a level-set curve, which is evolved to lie at the interior boundary of the nuclei surrounding the gland structure. Features are calculated from the boundaries that characterize the morphology of the lumen and the gland regions, including area overlap ratio, distance ratio, standard deviation and variance of distance, perimeter ratio, compactness, smoothness, and area. The feature space is reduced using a manifold learning scheme (Graph Embedding) that is used to embed objects that are adjacent to each other in the high dimensional feature space into a lower dimensional embedding space. Objects embedded in this low dimensional embedding space are then classified via a support vector machine (SVM) classifier as belonging to Gleason grade 3, grade 4 cancer, or benign epithelium. We evaluate the efficacy of the automated segmentation algorithm by comparing the classification accuracy obtained using the automated segmentation scheme to the accuracy obtained via a user assisted segmentation scheme. Using the automated scheme, the system achieves accuracies of 86.35% when distinguishing Gleason grade 3 from benign epithelium, 92.90% distinguishing grade 4 from benign epithelium, and 95.19% distinguishing between Gleason grades 3 and 4. The manual scheme returns accuracies of 95.14%, 95.35%, and 80.76% for the respective classification tasks, indicating that the automated segmentation algorithm and the manual scheme are comparable in terms of achieving the overall objective of grade classification.