Enhancing cervical squamous cell carcinoma diagnosis under distribution shift via domain-adaptive representation learning

Abstract

Summary Automated whole-slide image (WSI) analysis for cervical squamous cell carcinoma (CSCC) is challenged by limited pathological specificity in general models and poor cross-center generalization. We propose CSCC-PAG (pre-training, adaptation, generalization), a framework integrating domain-specific pre-training with self-supervised domain adaptation. Utilizing 8,041 CSCC WSIs, we applied the self-supervised pre-training for CSCC (SPC) feature extractor, which significantly outperforms pan-cancer baselines in the typing, grading, and segmentation tasks. To address domain shift, we developed representative sample domain adaptation (RSDA), combining masked autoencoders (MAEs) with a k-means buffer strategy for label-free adaptation. The results demonstrate that SPC captures fine-grained CSCC morphology, while RSDA effectively recovers performance degradation across centers without requiring target-domain annotations. This study provides a comprehensive solution for automated CSCC diagnosis, covering the entire pipeline from large-scale pre-training to few-shot generalized deployment, demonstrating its immense potential for real-world clinical implementation.