This paper proposes INSIGHT, a novel weakly supervised learning-based aggregator for analyzing large-scale medical images (3D CT scans and electron microscope images). To address the inherent shortcomings of existing methods, such as failure to localize small but important details and reliance on post-visualization techniques, INSIGHT integrates heatmap generation with inductive bias. Starting from pre-trained feature maps, it leverages a detection module with a small convolutional kernel and a context module with a large receptive field to capture fine details and suppress localized false positives. The resulting internal heatmap highlights diagnostically important regions. On CT and WSI benchmarks, INSIGHT achieves state-of-the-art classification performance and high weakly supervised learning-based semantic segmentation performance.
