Scanner- and stain-robust contrastive fine-tuning with matched-region positives
Post-train a pathology foundation model with a model-agnostic contrastive objective that explicitly reduces scanner- and stain-specific sensitivity by aligning embeddings of spatially corresponding tissue regions across different acquisition conditions (scanners and stains) while using other regions from the same condition as negatives. Implemented with parameter-efficient LoRA adapters trained on matched multi-scanner and multi-stain whole-slide images.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
Histopathology representations can encode scanner- and site-specific biases, which can produce generalization gaps. Contrastively tuning the model to associate corresponding regions across conditions aims to yield more robust, transferable representations for downstream tasks.
What changed
Introduced a contrastive post-training framework (SPECTRA) that adds scanner/stain robustness to existing pathology foundation models by (1) using matched multi-acquisition whole-slide data, (2) defining positives as spatially corresponding regions across scanners/stains, and (3) defining negatives as other locations within the same acquisition condition; training is performed via LoRA adapters.
Where it helps
- Improving generalization of pathology foundation models across scanners and stains
- Downstream pathology prediction tasks that suffer from site/acquisition shift
- Aligning embeddings for retrieval across acquisition conditions
- Parameter-efficient fine-tuning of existing foundation backbones for robustness
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
Reported gains were task-dependent (e.g., exceptions for segmentation and calibration), and the approach relies on availability of matched multi-scanner/multi-stain data with spatial correspondence.
Related World Models papers
Sources
- Contrastive fine-tuning against stain and scanner bias improves robustness and downstream performance across pathology foundation models ↗
arXiv · primary · Checked 2026-11-10
