Learning histopathological regions of interest by fusing bottom-up and top-down information

A virtual microscope is a system aimed to interact with digitized Whole Slide Images by displaying different Regions of Interest at any desired magnification and quality. Overall, an expert affords a diagnosis by examining a minimal number of regions, usually those with more information. Identification of such regions can improve the interaction process and optimize the examination workflow. This article presents a novel approach that learns relevant regions by integrating two information sources: bottom-up, computed from a Visual Attention Model that is structured as a graph, and top-down, captured from pathologists´ navigations that tune the information stored in the graph. The method was assessed against the pure VAM by comparing the percentage of requested information that was stored in a cache space during actual navigations. This method outperformed the percentage of available cache obtained with a simple Visual Attention Model model in about 8% after the first learning iteration.