DaNet: Domain-adaptive white blood cell classification through synthetic augmentation and cross-domain feature alignment

Abstract

Background and Objective:

Automated classification of white blood cells (WBCs) plays a vital role in improving clinical diagnostics and disease monitoring. However, current methods frequently face challenges with generalization, as they depend on training and testing data drawn from the same distribution. This limitation hinders their effectiveness in real-world clinical settings.

Methods:

We introduce DaNet, an innovative domain-adaptive method for classifying WBCs that leverages domain generalization techniques. DaNet comprises two key components: Balanced Multi-Domain Mixup (BMDM) and Data Distribution Alignment (DDA). BMDM serves as a data augmentation technique specifically designed to address the high similarity and class imbalance inherent in WBC datasets. By generating synthetic data that captures more distinctive and discriminative features, BMDM enhances the model’s ability to learn robust representations. DDA further aligns these features across different domains, enabling the model to learn domain-invariant characteristics.

Novelty:

The proposed BMDM facilitates a more continuous latent space for domain-invariant features across multiple source domains, which effectively alleviates the alignment challenges commonly encountered by DDA-based methods in WBC image classification tasks, particularly in the presence of subtle inter-class differences and class imbalance.

Results:

Extensive experiments demonstrate that DaNet performs reliably in WBCs classification tasks, particularly excelling in cross-domain generalization. The method shows solid performance across various blood cell classification tasks, indicating its effectiveness. Its innovative approach to data augmentation and domain alignment enhances the model’s robustness and generalizability.