Agent Aggregator with Mask Denoise Mechanism for Histopathology Whole Slide Image Analysis

Abstract

Histopathology analysis is the gold standard for medical diagnosis. Accurate\nclassification of whole slide images (WSIs) and region-of-interests (ROIs)\nlocalization can assist pathologists in diagnosis. The gigapixel resolution of\nWSI and the absence of fine-grained annotations make direct classification and\nanalysis challenging. In weakly supervised learning, multiple instance learning\n(MIL) presents a promising approach for WSI classification. The prevailing\nstrategy is to use attention mechanisms to measure instance importance for\nclassification. However, attention mechanisms fail to capture inter-instance\ninformation, and self-attention causes quadratic computational complexity. To\naddress these challenges, we propose AMD-MIL, an agent aggregator with a mask\ndenoise mechanism. The agent token acts as an intermediate variable between the\nquery and key for computing instance importance. Mask and denoising matrices,\nmapped from agents-aggregated value, dynamically mask low-contribution\nrepresentations and eliminate noise. AMD-MIL achieves better attention\nallocation by adjusting feature representations, capturing micro-metastases in\ncancer, and improving interpretability. Extensive experiments on CAMELYON-16,\nCAMELYON-17, TCGA-KIDNEY, and TCGA-LUNG show AMD-MIL's superiority over\nstate-of-the-art methods.\n