Current Pathology Foundation Models are unrobust to Medical Center Differences (2501.18055v2)

Abstract: Pathology Foundation Models (FMs) hold great promise for healthcare. Before they can be used in clinical practice, it is essential to ensure they are robust to variations between medical centers. We measure whether pathology FMs focus on biological features like tissue and cancer type, or on the well known confounding medical center signatures introduced by staining procedure and other differences. We introduce the Robustness Index. This novel robustness metric reflects to what degree biological features dominate confounding features. Ten current publicly available pathology FMs are evaluated. We find that all current pathology foundation models evaluated represent the medical center to a strong degree. Significant differences in the robustness index are observed. Only one model so far has a robustness index greater than one, meaning biological features dominate confounding features, but only slightly. A quantitative approach to measure the influence of medical center differences on FM-based prediction performance is described. We analyze the impact of unrobustness on classification performance of downstream models, and find that cancer-type classification errors are not random, but specifically attributable to same-center confounders: images of other classes from the same medical center. We visualize FM embedding spaces, and find these are more strongly organized by medical centers than by biological factors. As a consequence, the medical center of origin is predicted more accurately than the tissue source and cancer type. The robustness index introduced here is provided with the aim of advancing progress towards clinical adoption of robust and reliable pathology FMs.

• The paper introduces the Robustness Index to quantify how pathology models are biased by medical center variations.
• It evaluates ten models using cosine similarity in embedding spaces, revealing only one favored biological features.
• The findings stress the need for refined AI methods to mitigate confounding staining and center artifacts in pathology.

Robustness of Pathology Foundation Models Against Medical Center Variability

The paper “Current Pathology Foundation Models are unrobust to Medical Center Differences” addresses a critical challenge in the adoption of foundation models (FMs) in clinical pathology: the influence of confounding factors associated with medical center variations. This work empirically evaluates the robustness of pathology FMs and introduces methodologies to quantify the extent to which these models focus on biological features as opposed to confounding features, such as differences in staining procedures across medical centers.

The primary contribution of this paper is the introduction of the Robustness Index, a novel metric designed to evaluate the degree to which foundation models prioritize biological features over confounding ones. This index provides a quantitative measure of model robustness by assessing the neighborhood structure of the embedding spaces generated by the FMs. Specifically, the Robustness Index is defined as the ratio of the number of nearest neighbors with the same biological class to those with the same medical center, across samples in a dataset. The computation relies on cosine similarity within the embedding space to determine proximity.

The paper evaluated ten publicly available pathology FMs, revealing significant variations in robustness. Remarkably, only one model exhibited a Robustness Index greater than one, indicating a slight dominance of biological features over confounding medical center features in its embedding space organization. This suggests that the current path of FM development requires further optimization to ensure model outputs are more reflective of biological condition rather than influenced by site-specific artifacts.

In examining the effects of medical center differences, the paper uncovers pronounced clustering behavior in the 2D visualizations of FM embeddings organized using t-SNE. The analysis reveals a stronger organization by medical centers compared to biological classes, implying that medical center-related features inadvertently overshadow biological information, a finding corroborated by high medical center prediction accuracy compared to biological classification tasks.

The implications of these findings stress the need for enhanced approaches to understand and mitigate confounding signatures in pathology FMs, which is crucial to their reliable integration into clinical practice. The paper suggests that a model's clinical utility hinges on its ability to balance biological feature extraction with robustness against confounders—a vivid illustration of the broader challenges in developing and deploying AI in healthcare.

Future research avenues should explore strategies to disentangle confounding influences from genuine biological signals, potentially leveraging advanced self-supervised learning paradigms or more sophisticated data augmentation and normalization techniques. The paper also invites further studies on enhancement of the FM embedding spaces to facilitate bias reduction without sacrificing diagnostic accuracy.

In conclusion, this paper delivers significant insights into the robustness challenges facing current pathology FMs, setting groundwork for future research directions that could drive the successful translation of these models from research tools to clinically impactful technologies.