DELTA: Training-free and label-free correction of pathology foundation model embeddings through paired counterfactual stain-appearance interventions
DELTA: Training-free and label-free correction of pathology foundation model embeddings through paired counterfactual stain-appearance interventions
Choi, Y. H.; Kang, J.; Shin, S.-H.; Kim, C.
AbstractStaining and scanning variation impairs pathology foundation model (FM) generalization. We propose DELTA, a training-free, label-free method for post-hoc correction of FM embeddings. DELTA estimates recurrent directions from paired embedding differences between original patches and counterfactual stain-appearance variants preserving spatial structure, then attenuates aligned components through a fixed projection. Estimation from 4% of eligible patches per source dataset required no separately acquired matched images. Operators directly correct stored embeddings. On PathoROB, DELTA improved patch-level representation robustness over raw embeddings in all 20 FMs. Mean classification gains over established correction methods persisted after patch aggregation. DELTA provided additional gains after stain normalization. Whole-slide analyses showed improved in-distribution survival discrimination in TCGA-BRCA and normative cancer screening in a clinical cohort. Operators from TCGA-BRCA reduced embedding variation in the PLISM dataset across 13 staining conditions and 7 scanners without refitting. With projection rank and center matched, DELTA outperformed projections from mismatched pairs in retrieval and tissue-category concordance. Corrected embeddings retained information for spatial gene-expression prediction in the STP-BENCH datasets, with gains for tumor microenvironment markers in selected cohorts.