Swin Transformer and Momentum Contrast (MoCo) in Leukemia Diagnostics: A New Paradigm in AI-Driven Blood Cell Cancer Classification

Abstract

Acute Lymphoblastic Leukemia (ALL) is a rapidly progressing blood cancer requiring timely and accurate diagnosis for effective treatment. Automated image-based diagnostic models offer a potential solution but often lack robustness in clinical applications. This study aims to develop a robust, automated classification model for ALL subtypes using peripheral blood smear images, employing advanced feature extraction through the Swin Transformer framework. The proposed model utilizes the Swin Transformer architecture combined with Momentum Contrast (MoCo) for contrastive learning. Swin Transformer’s patch-based embedding and multi-level attention capture local and global cellular features, enhancing feature discrimination across ALL subtypes. MoCo further improves feature extraction by generating distinct embeddings, reducing overlap between cell types. The model achieved an overall classification accuracy of 92.5%, with a precision of 90.3%, recall of 91.1%, and F1-score of 90.7% across four classes (Benign, Malignant Early Pre-B, Malignant Pre-B, and Malignant Pro-B). Class-specific performance metrics indicate that Malignant Pre-B achieved the highest F1-score of 92.4%. The MoCo framework demonstrated a contrastive loss reduction from 0.5 to 0.097 for benign cells, significantly enhancing feature discrimination. Additionally, an ablation study showed that omitting the dynamic queue reduced accuracy by 5%, highlighting its importance for robust feature learning. This approach can be extended to other hematologic malignancies, with potential enhancements using more extensive datasets and integration with real-time diagnostic workflows in clinical settings to support precision medicine.