Abstract
White blood cell (WBC) classification is a clinically critical task in hematology, underpinning the diagnosis of blood malignancies, including leukemia, as well as infections and immune disorders. Accurate automated WBC differential counting is essential for early detection of conditions such as chronic lymphocytic leukemia (CLL) and acute myeloid leukemia (AML), where abnormal leukocyte counts and morphology are primary diagnostic indicators. Manual differential counting is time-consuming, subjective, and prone to inter-observer variability. This paper proposes a weighted ensemble framework combining three lightweight pretrained convolutional neural networks (CNNs): EfficientNet-B0, MobileNetV3-Small, and MobileNetV3-Large for automated four-class WBC classification on Paul Mooney’s Blood Cell Images dataset (BCCD, Kaggle). A standardized preprocessing pipeline was employed to enhance nuclear contrast and normalize photometric variability. Each backbone was finetuned under a two-phase transfer-learning protocol with AdamW optimization and cosine annealing scheduling. Model evaluation was conducted via stratified 3-fold cross-validation, and ensemble predictions were aggregated using accuracy-weighted soft voting. The proposed ensemble achieved a classification accuracy of 98.57%, a weighted F1-score of 0.9857, a Matthews Correlation Coefficient (MCC) of 0.9810, and a mean AUC of 0.9994. These results demonstrate the effectiveness of ensemble-based transfer learning combined with targeted preprocessing for robust, well-calibrated WBC classification, with direct applicability to clinical laboratory automation.