[Paper Review] Imbalanced Domain Generalization for Robust Single Cell Classification in Hematological Cytomorphology

Rao Muhammad Umer, Armin Gruber+3 more|arXiv (Cornell University)|Mar 14, 2023

Digital Imaging for Blood Diseases4 citations

TL;DR

This paper proposes a novel imbalanced domain generalization (IDG) framework for robust single-cell white blood cell (WBC) classification in hematological cytomorphology, combining domain-invariant representation learning with class-balanced loss optimization to address cross-domain data imbalance and distribution shifts. The method achieves state-of-the-art F1-macro scores on unseen test domains, significantly improving minority class recognition—particularly for rare cell types like lymphocyte_atypical—outperforming existing SOTA methods in out-of-distribution generalization.

ABSTRACT

Accurate morphological classification of white blood cells (WBCs) is an important step in the diagnosis of leukemia, a disease in which nonfunctional blast cells accumulate in the bone marrow. Recently, deep convolutional neural networks (CNNs) have been successfully used to classify leukocytes by training them on single-cell images from a specific domain. Most CNN models assume that the distributions of the training and test data are similar, i.e., the data are independently and identically distributed. Therefore, they are not robust to different staining procedures, magnifications, resolutions, scanners, or imaging protocols, as well as variations in clinical centers or patient cohorts. In addition, domain-specific data imbalances affect the generalization performance of classifiers. Here, we train a robust CNN for WBC classification by addressing cross-domain data imbalance and domain shifts. To this end, we use two loss functions and demonstrate their effectiveness in out-of-distribution (OOD) generalization. Our approach achieves the best F1 macro score compared to other existing methods and is able to consider rare cell types. This is the first demonstration of imbalanced domain generalization in hematological cytomorphology and paves the way for robust single cell classification methods for the application in laboratories and clinics.

Motivation & Objective

• To address the challenge of cross-domain data imbalance and domain shift in single-cell WBC classification across multiple clinical imaging sources.
• To improve model generalization to out-of-distribution (OOD) test domains, especially for under-represented and rare cell types.
• To develop a deep learning framework that maintains high performance on minority classes despite imbalanced training data across distinct hospitals or imaging protocols.
• To enable robust, clinically reliable AI-based classification in real-world hematological diagnostics where data distributions vary significantly.

Proposed method

• The method employs a dual-loss training strategy: a standard cross-entropy loss for classification and a domain-invariant contrastive loss in the latent feature space to reduce domain shift.
• It uses a Mahalanobis-like distance metric, $ d(\mathbf{z}, \{\boldsymbol{\psi}_{d,c}, \boldsymbol{\Sigma}_{d,c}\}) = \frac{1}{N_{d_i,c_i}} \sqrt{(\mathbf{z}-\boldsymbol{\psi}_{d,c})^\top \boldsymbol{\Sigma}_{d,c}^{-1}(\mathbf{z}-\boldsymbol{\psi}_{d,c})} $, to model class-specific feature distributions per domain.
• The model is trained with both coupled and decoupled optimization of the encoder and classifier heads, with the latter using class-balanced sampling to mitigate data imbalance.
• A ResNet50 backbone pre-trained on ImageNet is fine-tuned using five-fold cross-validation on source domains (Matek_19, Acevedo_20) and evaluated on an unseen target domain (INT_20).
• The framework explicitly handles missing classes in certain domains by learning domain-invariant representations that generalize across data-scarce scenarios.

Experimental results

Research questions

• RQ1 Can a deep learning model achieve robust out-of-distribution generalization in single-cell WBC classification when training data exhibit significant cross-domain data imbalance?
• RQ2 How does joint optimization of representation learning and classification loss improve performance on rare and under-represented WBC subtypes across diverse imaging sources?
• RQ3 To what extent does domain-invariant feature learning mitigate performance degradation due to domain shifts in clinical imaging protocols?
• RQ4 Does decoupled training with class-balanced sampling improve macro-F1 scores on minority classes compared to standard domain generalization methods?

Key findings

• The proposed IDG method achieves the highest F1-macro score of 0.78 ± 0.05 on the unseen INT_20 test set, outperforming ERM (0.40 ± 0.05), DANN (0.35 ± 0.06), and CORAL (0.43 ± 0.03).
• On the source domain validation sets (Acevedo_20 and Matek_19), the method achieves an F1-macro score of 0.78 ± 0.05, surpassing the next-best method (CORAL) by 0.02 points.
• The model significantly improves recognition of minority classes—such as lymphocyte_atypical—compared to standard ERM, as shown in confusion matrices where ERM misses many such samples.
• The decoupled training variant (Ours +) maintains competitive performance despite higher variance, indicating robustness under data imbalance.
• Domain generalization methods like DANN underperform when data are imbalanced, highlighting the necessity of integrating class balancing into DG frameworks.
• The use of a Mahalanobis-like distance in latent space enables better modeling of class-specific distributions across domains, contributing to improved OOD generalization.