Dice Question Streamline Icon: https://streamlinehq.com

Direct effect of adipose tissue on ABMIL predictions

Ascertain the degree to which adipose tissue directly affects predictions of attention-based multiple instance learning (ABMIL) models for breast cancer metastasis detection, quantitatively measuring whether adipose patches cause false negatives across specimens and models.

Information Square Streamline Icon: https://streamlinehq.com

Background

Attention maps highlighted adipose tissue in a misclassified specimen, but attention does not quantify causal influence on predictions. The authors explicitly state uncertainty about adipose tissue’s direct effect.

HIPPO is used to test this by removing or adding adipose patches, suggesting a broader need to systematically quantify how non-tumor tissue components affect ABMIL metastasis decisions.

References

Since attention maps only provide a qualitative visualization of regions in an image that the ABMIL models consider important, it is unclear to what extent adipose tissue directly affects model predictions.

— Explainable AI for computational pathology identifies model limitations and tissue biomarkers (2409.03080 - Kaczmarzyk et al., 4 Sep 2024) in Results — Subsection “Adipose tissue can cause false negative metastasis detections”