Robustifying Pathology Foundation Models via Fine-Tuning

This paper introduces a fine-tuning recipe for improving pathology foundation models under scanner and staining variability. Evaluated across ten different foundation models, the method reportedly improved both robustness and downstream performance without an observed trade-off. The average PathoROB robustness index increased from 0.72 to 0.87, while combined performance across Patho-Bench, HEST, and THUNDER improved by 43%. Reported peak gains reached 72% in robustness for Phikon-v2 and 76% in performance for Midnight-12k. Fine-tuned versions of Phikon-v2 and Midnight-12k are released as Phaet and Mascaret.