The problem — A gigapixel whole-slide image (WSI) cannot be fed to a classifier in one pass. It must be cut into thousands of tiles, yet only a slide-level diagnosis label is typically available — no tile has a ground-truth annotation. Standard supervised learning has nothing to train on at the patch level; a model that ignores this structure will average signal across irrelevant tissue.
What it is / how it works — ABMIL formalises this as a multiple instance learning problem. The slide is the bag; its tiles are instances. A pre-trained encoder (historically ResNet, now often a pathology foundation model such as UNI) maps each tile to a feature vector. ABMIL then learns a small attention network that produces a scalar weight per tile. The weighted mean of tile features becomes the slide-level representation, which a final classifier reads off. Crucially, the attention weights are interpretable: high-weight tiles are the model's evidence for its prediction, providing a form of spatial localisation without pixel-level supervision. A gated attention variant adds a second sigmoid branch for sharper discrimination.
Where it fits — ABMIL is the standard aggregation head for WSI classification tasks — survival prediction, subtype identification, biomarker detection — across the digital pathology model ecosystem. It is not a feature extractor; it assumes good tile representations and learns only how to pool them.
Where it breaks — Attention pooling assumes tiles are exchangeable given the bag label. Spatial context between adjacent tiles is discarded; a tumour at a margin may receive low attention if the spatial relationship is diagnostically critical. Attention maps can be misleading: high weight is correlation with the training label, not causal tissue identity. Evaluation on a single cohort inflates performance — scanner-to-scanner and stain-protocol shifts silently degrade attention alignment.