Two-stage CNN-based framework for leukocytes classification

Abstract

Leukocytes are pivotal markers in health, crucial for diagnosing diseases like malaria and viral infections. Peripheral blood smear tests provide pathologists with vital insights into various medical conditions. Manual leukocyte counting is challenging and error-prone due to their complex structure. Accurate segmentation and classification of leukocytes remain challenging, impacting both accuracy and efficiency in blood microscopic image analysis. To overcome these limitations, we propose a robust two-stage CNN framework that integrates YOLOv8 for precise segmentation and MobileNetV3 for effective classification. Initially, WBCs are segmented using YOLOv8m-seg, extracting ROIs for subsequent analysis. Then, features from segmented ROIs are used to train MobileNetV3, classifying WBCs into lymphocytes, monocytes, basophils, eosinophils, and neutrophils. This framework significantly advances leukocyte categorization, enhancing diagnostic performance and patient outcomes. The proposed technique achieved impressive accuracy rates of 99.56 %, 99.19 % and 98.89 % during segmentation and 99.28 %, 99.63 % and 98.49 % during classification on Raabin-WBC, PBC and LISC datasets, respectively, outperforming state-of-the-art methods.