Domain Adaptation for Breast Cancer Computational Pathology: Evaluating Feature-Level and Pixel-Level Approaches Under Leave-One-Domain-Out Framework

Abstract

Deep learning models for computational pathology suffer substantial performance degradation when deployed across institutions with different scanners, staining protocols, and patient populations. This thesis investigates whether domain adaptation can improve cross-institutional performance of a multiple instance learning classifier for estrogen receptor (ER) status prediction in breast cancer, evaluated across four geographically diverse whole-slide image datasets (TCGA-BRCA, Zanmi Lasante Haiti, BCNB, and PostNAT-BRCA) under a Leave-One-Domain-Out framework. Three aims were pursued. Aim 1 established a CLAM-CONCH baseline using a frozen foundation model encoder. The baseline achieves held-out AUC exceeding 0.76 across all four folds, with cross-domain gaps of 13–17 percentage points for the three large-cohort folds. Feature space analysis confirms strong domain-driven clustering despite histopathology-specific pretraining, motivating explicit adaptation. Aim 2 evaluated feature-level adaptation via Deep CORAL, integrated as a covariance alignment loss within CLAM training. Slide-level CORAL produced the largest improvement for TCGA (+4.54 pp), the domain with the greatest distributional shift, while providing marginal gains for Haiti and BCNB. Tile-level CORAL failed across all domains, attributable to the architectural bottleneck of a single linear layer between the frozen encoder and the attention pooling mechanism. Aim 3 evaluated pixel-level adaptation via CyCADA: a CycleGAN augmented with a semantic consistency loss using frozen CONCH features to constrain content preservation during translation. CycleGAN-only adaptation recovered approximately 51% of the TCGA domain gap (+8.59 pp). Adding semantic consistency achieved the best result for BCNB (+6.68 pp) and was the only pixel-level variant to improve on baseline for Haiti (+0.62 pp), preventing the degradation produced by unconstrained translation. The central finding is that adaptation effectiveness depends on the interaction between method and domain shift type. Pixel-level style transfer suits high-resource domains with large appearance-driven shift; feature-level alignment suits distributional shift; and neither approach adequately addresses acquisition-process degradation of diagnostic signal quality, as observed in Haiti. Haiti's resistance to all tested methods identifies a failure mode that standard benchmarks conceal and motivates adaptation strategies designed for resource-constrained clinical environments.