[논문 리뷰] Robust White Blood Cell Classification with Stain-Normalized Decoupled Learning and Ensembling

Luu Le, Hong Cao외 3명|arXiv (Cornell University)|2026. 03. 02.

Digital Imaging for Blood Diseases인용 수 0

한 줄 요약

본 논문은 도메인 시프트 하에서 긴 꼬리 WBC 분류를 다루기 위해 stain-normalized, decoupled 두-stage 학습 프레임워크를 제시하고, 테스트 시 증강 및 백본 앙상블을 도입하여 WBCBench 2026에서 최상위 성능을 달성한다.

ABSTRACT

White blood cell (WBC) classification is fundamental for hematology applications such as infection assessment, leukemia screening, and treatment monitoring. However, real-world WBC datasets present substantial appearance variations caused by staining and scanning conditions, as well as severe class imbalance in which common cell types dominate while rare but clinically important categories are underrepresented. To address these challenges, we propose a stain-normalized, decoupled training framework that first learns transferable representations using instance-balanced sampling, and then rebalances the classifier with class-aware sampling and a hybrid loss combining effective-number weighting and focal modulation. In inference stage, we further enhance robustness by ensembling various trained backbones with test-time augmentation. Our approach achieved the top rank on the leaderboard of the WBCBench 2026: Robust White Blood Cell Classification Challenge at ISBI 2026.

연구 동기 및 목표

• Address cross-domain variation from staining and scanning in WBC classification.
• Mitigate long-tailed class distribution to improve minority class performance.
• Develop a decoupled training pipeline separating representation learning from classifier rebalancing.
• Incorporate stain normalization, test-time augmentation, and model ensembling to boost robustness.

제안 방법

• Apply Macenko stain normalization to align stain appearance across domains.
• Use instance-balanced sampling for Stage 1 representation learning.
• Freeze backbone and retrain classifier with class-balanced sampling in Stage 2.
• Employ a hybrid loss combining effective-number re-weighting and focal modulation during Stage 2.
• In inference, perform test-time augmentation and ensemble multiple backbones (CNNs and ViT) to improve robustness.

Fig. 1 : Overview of our proposed framework for WBC classification

실험 결과

연구 질문

• RQ1 Can stain normalization and a decoupled training strategy improve cross-domain WBC classification under long-tailed distributions without target-domain labels?
• RQ2 How do class-balanced sampling and a hybrid loss affect minority class sensitivity and overall macro metrics?
• RQ3 Does combining CNN and transformer backbones with test-time augmentation yield better robustness than single-model approaches?

주요 결과

• Decoupled learning improves macro-balanced metrics, e.g., ResNet50 from MF1 65.8 to 70.4 and ResNet152 from MF1 68.7 to 70.6.
• Backbone Swin achieves highest single-model BAcc under decoupling (80.9) but lower MF1 and MP, indicating calibration differences across architectures.
• Ensembling backbones (R50+R152, R50+Swin, R152+Swin) yields further gains, with the full ensemble achieving MF1 74.2 and BAcc 77.1.
• Overall, class-aware rebalancing and hybrid loss enhance minority sensitivity with limited impact on majority classes.
• No target-domain labels are used for the final robustness gains, relying on decoupling, TTA, and ensembling.

Fig. 2 : Confusion matrix of our best model on WBCBench 2026 Challenge Benchmark.