Automated classification of pediatric acute lymphoblastic leukemia: A ResNet-50 deep learning approach

Abstract

Early detection of acute lymphoblastic leukemia (ALL) is crucial for improving survival outcomes in children. Manual diagnosis through microscopic examination is often time-consuming and subject to human error. This study presents an automated classification framework for pediatric ALL using a fine-tuned Residual Network (ResNet)-50 deep learning architecture. The model was trained and validated on 15,135 segmented blood smear images collected from 118 pediatric patients in the publicly available ALL IDB Version 2 dataset. Data augmentation and patient-wise splitting were applied to ensure model generalization and prevent data leakage. The fine-tuned ResNet-50 achieved a mean classification accuracy of 99.60%, with precision, recall, and F1-score of 99.45%, 99.40%, and 99.42%, respectively, outperforming baseline convolutional neural network models. Statistical validation (p < 0.0015) confirmed that these performance improvements are highly significant. This study highlights the potential of ResNet-50 for reliable, automated, and reproducible leukemia diagnosis, offering clinical decision support for early detection and treatment planning.