AI-Assisted Web Application for Leukocyte Abnormality Counting with YOLOv11 and Smartphone Microscopy

Abstract

Accurate and timely white blood cell (WBC) analysis is crucial for diagnosing hematological disorders, often requiring microscopic examination of peripheral blood smears (PBS). While manual counting by trained specialists is considered the gold standard, it is time-consuming and impractical in resource-limited settings. Automated cell counters can misclassify similitude or immature cells, hindering accurate diagnosis. To address these limitations, we propose an AI-powered web application that utilizes YOLOv11 with enhanced small object detection capabilities, enabled by integrating our C3k2-Conv blocks, an architecture inspired by C3k2. Our model, trained on eleven WBC classes and nucleated red blood cells (NRBCs), achieves an impressive mean average precision (mAP@0.5) of approximately 0.9000 on validation and unseen test sets, demonstrating a performance comparable to human specialists in identifying and quantifying WBCs. Furthermore, our research demonstrates that providing general practitioners and medical students with PBS images annotated by our AI model significantly improves their counting accuracy and reduces the time spent on manual counting. Our web application, Myelosoft, allows clinicians to upload smartphone-captured PBS images for rapid and automated analysis. The system provides comprehensive differential counts for 11 WBC classes, including atypical lymphocytes, band neutrophils, basophils, blasts, eosinophils, lymphocytes, metamyelocytes, monocytes, myelocytes, promyelocytes, and segmented neutrophils, as well as NRBCs. This real-time analysis facilitates timely diagnosis and treatment, potentially reducing risks associated with delayed interventions. Our approach offers a robust and accessible solution for improving hematologic treatment, especially in resource-constrained environments.