Interpretable WBC Analysis through Pathologist-Aligned Reasoning
Year: 2026
Type: article
Abstract: White blood cell (WBC) analysis, encompassing segmentation, morphological attribute analysis, and classification, is vital for diagnosing blood-related conditions. Pathologists follow a systematic workflow: identifying WBCs, analyzing structural features, and classifying them based on morphology. Existing systems, however, overlook their synergy and often address these tasks in isolation. Also, these systems are incapable of analyzing global contextual information (or a holistic understanding of cell morphology) alongside subtle structural variations (or localized features). Both these limitations restrict the performance of existing systems and hinder their interpretability. To address these limitations, we propose a system that synergizes segmentation, morphological attribute prediction (MoAP), and classification. It mirrors the pathologist’s diagnostic process for the synergization. Also, it proposes a novel hybrid architecture for simultaneously analyzing local and global information. Specifically, our architecture combines convolutional neural networks (CNNs) for fine-grained structural analysis with a novel cross-attention-based Transformer decoder that emphasizes global context. Experimental results on publicly available datasets demonstrate that our system outperforms state-of-the-art systems for WBC segmentation, MoAP, and WBC classification. Also, it shows that our system can be easily fine-tuned for effective cross-dataset evaluations.
DOI: 10.1145/3822598
Authors: Rajesh Kumar, Adit Srivastava, Puneet Gupta
Institution: Indian Institute of Technology Indore
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166030542