FOCI: A Lightweight Method for Rationale Highlighting in Frozen WSI-MIL
A recent study presents Finding Optimal Contextual Instances (FOCI), a streamlined rationale-readout layer designed for frozen whole-slide image (WSI) multiple instance learning (MIL) classifiers. This innovative approach seeks to derive slide-level predictions from a compact, output-consistent tile subset without the need for backbone retraining. FOCI employs model-output sufficiency and exclusion objectives on keep/drop tile subsets, assessed through an insertion-style Sequential Reveal Protocol (SRP) tailored for WSI-MIL, and summarized using the Selection Headroom Index (SHI). The study tackles the lack of clarity in attention scores as post-hoc explanations in WSI-MIL, indicating that elevated attention might indicate aggregation preference instead of a concise rationale. Experiments were performed across three WSI benchmarks and seven MIL backbones. This research is available on arXiv with the identifier 2605.12575.
Key facts
- FOCI is a lightweight rationale-readout layer for frozen WSI-MIL classifiers.
- It recovers slide-level predictions from a compact tile subset without retraining.
- Trained with model-output sufficiency and exclusion objectives.
- Evaluated using Sequential Reveal Protocol (SRP) adapted to WSI-MIL.
- Summarized by Selection Headroom Index (SHI).
- Addresses opacity of attention scores as post-hoc explanations.
- Experiments on three WSI benchmarks and seven MIL backbones.
- Published on arXiv:2605.12575.
Entities
Institutions
- arXiv