Spatial Blindness in Whole-Slide Multiple Instance Learning

In three linesWhole-slide MIL models suffer from 'spatial blindness': accurate predictions but ignoring tissue architecture. ResTopoMIL fixes this by separating appearance statistics (prototype histogram) from spatial relations (graph branch with permutation constraint). Improvements across 9 WSI benchmarks with 1.15M parameters.

Summary generated by Claude — human-verified