Gland-based Prostate Tissue Image Classification

Here, we address the problem of classifying the prostate tissue images as normal, Gleason grade 3 and grade 4, by using information computed from glandular regions in the image this classification determines the malignancy of cancer and is of great diagnostic importance. For this task, state-of-the-art methods rely on segmenting and extracting features from gland regions containing strong lumen. Our novelty is in our treatment of gland regions with weakly detectable lumen; in such regions, we locate the epithelial nuclei, combine them into patches based on spatial density constraints, and compute their features to improve the classification. To evaluate the proposed method, we use a database of 299 images at 20× magnification. The classification accuracies for our method are 95.9% for normal v. grade 3, 95.9% for normal v. grade 4, 84.7% for grade 3 v. grade 4 and 86.4% for three-class classification. These results outperform published methods in the literature.