Spatial Blindness in Whole-Slide Multiple Instance Learning
For computational pathology researchers, this work exposes a critical failure mode in MIL models and offers a simple fix, though the problem is domain-specific.
The paper identifies that many whole-slide MIL models, despite being labeled as context-aware, are actually spatially blind—their predictions remain nearly unchanged when patch coordinates are permuted. The authors propose ResTopoMIL, which trains a permutation-invariant prototype histogram first, then freezes it while a lightweight graph branch learns spatial residuals, achieving improved classification and survival prediction across 9 benchmarks with 1.15M parameters.
Whole-slide MIL models are often called context-aware once graphs, Transform ers, or state-space modules are placed above patch embeddings. We show that this label can be deceptive. On pathology tasks where tissue architecture is part of the diagnostic signal, several strong MIL baselines retain nearly unchanged slide level AUC after patch coordinates are permuted. Their predictions are accurate, but largely compositional. We refer to this failure mode as spatial blindness. Our explanation is optimization-based: dense appearance statistics are learned early under slide-level supervision, leaving weak gradients for sparse spatial relations. ResTopoMIL addresses the issue by first fitting a permutation-invariant prototype histogram and then freezing it while a lightweight graph branch learns the residual under a coordinate-shuffling constraint. The architecture is simple by design; the intervention is in how the spatial branch is trained. Across 9 public WSI bench marks, ResTopoMIL improves classification and survival prediction with 1.15M parameters, restores sensitivity to coordinate perturbation, and gives stronger lo calization evidence on CAMELYON-16.