GHMS-CycleGAN: Graph-Based Hierarchical Multi-stain CycleGAN for Stain Normalization and Classification in Digital Pathology.

The emergence of digital pathology (DP) has provided numerous computer-aided diagnosis (CAD) opportunities and tools based primarily on deep learning (DL) approaches. However, staining variations and distortions caused by DP scanners during image digitization pose challenges and cause performance degradation in DL-based CAD systems. In this work, we introduce novel methods of stain and scan normalization for histopathology, red blood cell (RBC), and white blood cell (WBC) image classification. In particular, we propose a novel framework for multi-stain and color normalization using hierarchical multi-stain cycle-consistent generative adversarial networks (HMS-CycleGAN) and graph-based HMS-CycleGAN (GHMS-CycleGAN). Our method effectively learns image normalization mappings by exploiting underlying hierarchical and network structures in DP images. The proposed framework was validated on histopathology, RBC, and WBC image datasets collected from different scanners and institutions to assess the impact of the proposed CycleGAN-based normalization methods on the downstream task of image classification. The results show that the HMS-CycleGAN normalizer can generate significantly more robust and consistent normalization across diverse staining variations compared to state-of-the-art methods. Also, our normalization methods consistently lead to superior classification performance compared to classifiers trained without image normalization and those using three state-of-the-art methods. HMS-CycleGAN achieved the best classification accuracy (84.05%) on the Wilds Camelyon dataset and the best F1-score for WBC classification (51.06%), while both HMS-CycleGAN and GHMS-CycleGAN excelled in cross-scanner RBC image classification, reaching F1-scores between 80 and 91%. The experimental results highlight the importance of image normalization in building robust and generalizable DL-based models for improved downstream image classification in DP.