Content-based Tissue Region Retrieval in Prostate Histopathology

In this paper, we address the tissue region retrieval problem in prostate histopathology: we search for tissue regions visually similar to a query region from a database of prostate tissue slide images. To achieve this goal, a gland-based method to compute the similarity between two tissue regions is adopted, in which we first need to segment glands and extract their features. The region similarity is computed using information about gland density, gland structure and gland area of the two regions. To evaluate the retrieval result, we assign the tissue regions into four different categories: background, normal, Gleason grade 3 and Gleason grade 4. By combining the gland-based method with a bag-ofwords-based method, we obtain almost 7 relevant regions (regions with the same categories as the query) in the top 10 retrieved regions, which is a better result than popular methods in medical image retrieval literature.