Sufficiency of tumor regions for ABMIL metastasis detection

Determine whether metastatic tumor regions alone are sufficient to drive positive predictions in attention-based multiple instance learning (ABMIL) models for breast cancer metastasis detection, for example by evaluating predictions when only tumor patches are present or when tumor patches are added to normal slides.

Background

Accurate metastasis detection requires sensitivity to even small tumor regions. The authors explicitly note uncertainty regarding whether tumor-only inputs suffice for positive predictions in ABMIL models.

Using HIPPO on CAMELYON16, they construct counterfactuals to test sufficiency, but the open question centers on establishing sufficiency across different model architectures and metastasis sizes.