Influence and necessity of tumor regions in ABMIL metastasis detection

Characterize the extent to which metastatic tumor regions influence predictions of attention-based multiple instance learning (ABMIL) models for breast cancer metastasis detection on whole slide images, including whether positive predictions depend on the presence of tumor patches.

Background

Although metastasis detection is driven by tumor cells, ABMIL models can learn shortcut features or rely on extratumoral tissue. The authors explicitly state uncertainty about how much tumor regions influence ABMIL metastasis predictions.

The paper uses HIPPO with expert tumor annotations in CAMELYON16 to test necessity by removing tumor patches, but the stated uncertainty frames the broader problem of rigorously characterizing reliance on tumor regions across models and contexts.

References

The degree to which tumor regions influence ABMIL models for metastasis detection remains unclear.

— Explainable AI for computational pathology identifies model limitations and tissue biomarkers  (2409.03080 - Kaczmarzyk et al., 2024) in Methods — Subsubsection “Testing the necessity of tumor regions”

We also observe that MaMA attends broadly across breast tissue rather than focusing on specific findings; we conjecture that this global conditioning allows the model to exploit shortcut features, which may partly explain its competitive linear probe performance in Table~\ref{tab:linearprobe} despite poor localization.

— Solving the Needle-in-a-Haystack Problem in Mammography Vision-Language Model with Differentiable Subset Sampling  (2609.03085 - Jeon et al., 2 Sep 2026) in Section 4.5, Results, subsection “Localization”