Robustifying pathology foundation models via fine-tuning

2026/07/24 by Alexandre Filiot, Oskar Thaeter, Benoit Schmauch +1
#cs.CV #cs.AI

paper · pdf

Abstract

Pathology foundation models (FMs) produce powerful tile-level representations which remain sensitive to scanner and staining variability, undermining deployment across laboratories. We develop a novel fine-tuning recipe that improves the robustness of pathology FMs to acquisition factors. Applied to ten different FMs, our fine-tuning strategy consistently improves robustness for every model as well as downstream performance, with no observed trade-off. On average, it raises the PathoROB robustness index by 23% (from 0.72 to 0.87) and increases the overall cross-benchmark performance by 43% on Patho-Bench, HEST and THUNDER combined, with individual gains reaching up to 72% in robustness (Phikon-v2) and 76% in performance (Midnight-12k). We publicly release the fine-tuned versions of Phikon-v2 (Phaet) and Midnight-12k (Mascaret) at https://huggingface.co/wearewaiv/models.

Citations

• Atlas 2 -- Foundation models for clinical deployment
• Pathryoshka: Compressing Pathology Foundation Models via Multi-Teacher Knowledge Distillation with Nested Embeddings
• Toward Clinically Grounded Foundation Models in Pathology
• DINOv3
• Pathology Foundation Models are Scanner Sensitive: Benchmark and Mitigation with Contrastive ScanGen Loss
• SCORPION: Addressing Scanner-Induced Variability in Histopathology
• THUNDER: Tile-level Histopathology image UNDERstanding benchmark
• MeDi: Metadata-Guided Diffusion Models for Mitigating Biases in Tumor Classification
• Training state-of-the-art pathology foundation models with orders of magnitude less data
• Beyond Diagnostic Performance: Revealing and Quantifying Ethical Risks in Pathology Foundation Models
• Accelerating Data Processing and Benchmarking of AI Models for Pathology
• Current Pathology Foundation Models are unrobust to Medical Center Differences
• Distilling foundation models for robust and efficient models in digital pathology
• Multimodal Whole Slide Foundation Model for Pathology
• Evaluating Computational Pathology Foundation Models for Prostate Cancer Grading under Distribution Shifts
• Phikon-v2, A large and public feature extractor for biomarker prediction
• Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology
• HEST-1k: A Dataset for Spatial Transcriptomics and Histology Image Analysis
• Domain Generalization in Computational Pathology: Survey and Guidelines
• ContriMix: Scalable stain color augmentation for domain generalization without domain labels in digital pathology
• DINOv2: Learning Robust Visual Features without Supervision
• iBOT: Image BERT Pre-Training with Online Tokenizer
• TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification
• Emerging Properties in Self-Supervised Vision Transformers
• Barlow Twins: Self-Supervised Learning via Redundancy Reduction
• Self-Supervision Closes the Gap Between Weak and Strong Supervision in Histology
• Self supervised contrastive learning for digital histopathology
• Improved Baselines with Momentum Contrastive Learning
• From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The CAMELYON17 Challenge
• Attention-based Deep Multiple Instance Learning
• Color transfer between images
• elastix: A Toolbox for Intensity-Based Medical Image Registration

Related

• Distilling foundation models for robust and efficient models in digital pathology
• Benchmarking Pathology Foundation Models: Adaptation Strategies and Scenarios
• Towards Robust Foundation Models for Digital Pathology
• Benchmarking pathology foundation models: Adaptation strategies and scenarios
• Training state-of-the-art pathology foundation models with orders of magnitude less data