Abstract
Stain variability and scanner-induced distortions remain major obstacles to reliable white blood cell (WBC) image analysis, limiting the generalization of deep learning models across domains. To overcome these challenges, we introduce THMS-CL-CycleGAN, a task-guided hierarchical multi-stain normalization framework with contrastive learning that jointly enhances visual consistency, structural fidelity, and classification robustness. The proposed model integrates two domain-specific classification discriminators to provide explicit task-aware supervision. This design effectively adds semantic features to ensure that normalized images retain discriminative, biologically meaningful cues. Furthermore, a contrastive loss is employed to ensure semantic consistency between source and target domains, while a hierarchical cycle-consistency loss preserves multi-scale contextual structures across domains. Experimental evaluations on two WBC datasets (Raabin (in-distribution) and Munich (out-of-distribution)) demonstrate that the proposed method achieves state-of-the-art performance in both stain normalization and downstream WBC classification. The THMS-CL-CycleGAN model achieves a macro-average F1-score of 87.47%, outperforming the no-normalization model by +22.9% and the best competing approach (HMS-CycleGAN) by +4.7%. Overall, these results demonstrate that combining task-aware supervision, contrastive learning, and hierarchical consistency constraints leads to notable improvements in domain adaptation and classification reliability. These findings position THMS-CL-CycleGAN as a powerful and reliable framework for multi-domain stain normalization and WBC image analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/HemaRAI_002.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=WTiKOMeKdr
BibTex
@InProceedings{ElmMoh_Taskguided_MICCAISAT2026,
author = { Elmanna, Mohamed AND Morsy, Ahmed AND Rushdi, Muhammad},
title = { { Task-guided Hierarchical Multi-stain CycleGAN with Contrastive Learning for Robust White Blood Cell Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17252},
month = {pending},
page = {pending}
}