TLDR: A new study introduces PathoROB, a benchmark to evaluate the robustness of AI Foundation Models (FMs) in digital pathology. It found that current FMs are often non-robust, learning non-biological technical features that lead to significant diagnostic errors. The research proposes and tests methods to make these models more robust without retraining, showing promising improvements but also highlighting the need for robustness as a core design principle in future AI development for healthcare.
Artificial intelligence (AI) models, particularly large-scale Foundation Models (FMs), are rapidly changing healthcare research and are on the verge of clinical use. However, a significant concern in digital pathology is their tendency to learn non-biological technical features, such as variations in surgical techniques, lab procedures, or scanner hardware. This can pose serious risks when these models are deployed in real clinical settings.
A recent research paper, “Towards Robust Foundation Models for Digital Pathology,” delves into this critical issue, presenting the first systematic investigation into the robustness of pathology FMs against these non-biological features. The study highlights that while FMs have shown remarkable capabilities in various diagnostic tasks, their reliability can be compromised by these technical variations, which do not reflect actual medical or biological tissue characteristics.
The researchers introduced a new framework called PathoROB, a robustness benchmark designed to systematically measure how well pathology foundation models perform when faced with non-biological variations across different medical centers. PathoROB includes three novel metrics, such as the “robustness index,” and utilizes four datasets covering 28 biological classes from 34 medical centers. This comprehensive approach allowed for a thorough evaluation of 20 different FMs.
The findings from PathoROB revealed widespread robustness deficits across all evaluated FMs, with significant differences in robustness between them. The study found that when FMs are not robust, their internal representations can lead to major diagnostic errors and clinical mistakes, which could prevent their safe adoption in healthcare. For instance, models might incorrectly classify a tumor based on the hospital it came from rather than its actual biological characteristics.
The paper also demonstrated the medical consequences of limited robustness. In supervised tasks, where FMs are used to develop specific diagnostic models, non-robust FMs can cause these downstream models to rely on technical shortcuts instead of true biological signals. This leads to poor generalization performance on new, unseen data. Similarly, applications like image clustering and diagnostic case search, which rely on the learned FM representations, were negatively impacted by these medical center signatures.
To address these issues, the researchers proposed a framework for “robustification” – making FMs more reliable without the need for costly retraining. They explored three main approaches: data robustification (removing technical signatures directly from images, like stain normalization), representation robustification (removing signatures after feature extraction), and training robustification (preventing downstream models from using these signatures during training). Methods like Reinhard stain normalization and ComBat batch correction significantly improved the robustness index, making the FM representations more biologically meaningful.
While these robustification methods considerably reduced the risk of errors, they did not entirely eliminate them. This suggests that biological and technical information can be deeply intertwined within FM representations, making complete disentanglement challenging. Nevertheless, the work emphasizes that evaluating robustness is crucial for validating pathology FMs before clinical use and that future FM development must integrate robustness as a core design principle.
Also Read:
- SmartPath-R1: A New AI System for Comprehensive Pathology Analysis
- Systematic Review Explores Multimodal Machine Learning for Cancer Survival Prediction
This research provides a valuable blueprint for assessing robustness across various biomedical domains, guiding efforts to create more robust, representative, and clinically deployable AI systems that prioritize biological information over technical artifacts. For more details, you can read the full paper here.
