Fine-tuning improves robustness of pathology foundation models
A novel fine-tuning approach significantly boosts the resilience of pathology foundation models (FMs) against variations in scanners and staining, which has been a major obstacle for clinical application. This method was tested on ten distinct FMs, yielding consistent enhancements in robustness and performance without any compromises. The PathoROB robustness index improved by an average of 23% (from 0.72 to 0.87), while overall performance across Patho-Bench, HEST, and THUNDER surged by 43%. Notable individual improvements included up to 72% in robustness for Phikon-v2 and 76% in performance for Midnight-12k. The fine-tuned Phikon-v2 (Phaet) and Midnight-12k (Mascaret) models are now available to the public. This research tackles a vital issue in computational pathology, where laboratory variability hampers model generalization.
Key facts
- Pathology foundation models are sensitive to scanner and staining variability.
- A novel fine-tuning recipe improves robustness to acquisition factors.
- Applied to ten different FMs, the strategy consistently improves robustness and downstream performance.
- PathoROB robustness index increased by 23% on average (from 0.72 to 0.87).
- Overall cross-benchmark performance increased by 43% on Patho-Bench, HEST, and THUNDER combined.
- Individual gains up to 72% in robustness (Phikon-v2) and 76% in performance (Midnight-12k).
- Fine-tuned versions of Phikon-v2 (Phaet) and Midnight-12k (Mascaret) are publicly released.
- The study addresses variability across laboratories limiting model deployment.
Entities
—