Multiple Instance Learning

MIST is evaluated as a plug-in projection-layer replacement across eight distinct MIL aggregators. Multiple instance learning has become the standard framework for whole-slide image analysis in computational pathology. Prior MIL research has invested heavily in the encoder and aggregation stages, leaving the projection layer comparatively underexplored. The canonical MIL pipeline decomposes a gigapixel slide into patches, encodes them with a frozen foundation model, applies a per-patch projection layer, and aggregates into a slide-level prediction.