SemaNorm: Stain Normalization via Semantic and Morphology Alignment for Robust Cross-Hospital Computational Pathology

Source abstract

Stain variability remains a major source of domain shift in computational pathology, yet existing normalization methods largely optimize color consistency without explicitly preserving morphology or diagnostic semantics. We propose SemaNorm, a self-supervised stain canonicalization stain normalizer to recover clean patches from strongly perturbed inputs, avoiding paired transfer, adversarial instability, and identity collapse. It further combines a frozen pathology foundation-model consistency loss with a differentiable nuclei-structure loss to preserve semantic and cellular organization. SemaNorm is evaluated across Camelyon17-WILDS and NCT-CRC-HE-100K using cross-hospital tumor classification, cell-graph morphology metrics, and semantic stability measured with an independent held-out foundation encoder. SemaNorm improves out-of-distribution accuracy by 5.9% over the strongest normalization baseline and by 13.5% over no normalization, while achieving the best morphology preservation. Independent foundation-model and cell-graph analyses further confirm improved semantic and structural stability. These results establish morphology- and semantics-aware optimization as a principled alternative to color-only stain normalization.