Whole Slide Image Domain Adaptation Tailored with Fisher Vector, Self Supervised Learning, and Novel Loss Function

Ravi Kant Gupta, Shounak Das, Shambhavi Shanker, Ardhendu Sekhar, Amit Sethi

Keywords: Domain Adaptation, Fisher Vector, Self-supervised, Whole Slide Image.

TL;DR: We propose a domain adaptation framework using self-supervised MoCoV3 features, k-means clustering, and Fisher Vector encoding, enhanced with PLMMD and MCC losses, achieving strong cross-domain HER2 classification on TCGA-BRCA and Warwick.

Abstract: Whole Slide Images (WSIs) present major challenges in computational pathology due to their high resolution, morphological diversity, and domain variability across institutions. These factors result in domain shifts that limit model generalization. We propose a domain adaptation framework that integrates self-supervised learning, clustering, and Fisher Vector encoding for robust WSI classification. Patch-level features are extracted using MoCoV3, clustered via k-means, and aggregated using Gaussian mixture-based Fisher Vectors to form compact slide-level representations. To align domains, we employ adversarial training enhanced with a tailored loss combining PLMMD and MCC. Evaluated on HER2 classification across TCGA-BRCA and Warwick datasets, our method consistently outperforms baselines, especially under label noise and domain shift, demonstrating the strength of combining self-supervised features with structured statistical encoding for cross-domain WSI analysis.

Submission Number: 26