Fine-Tuning Deep Learning Architectures for Leukemia Diagnosis from Blood Smears: A Comparative Study

Early and accurate diagnosis is pivotal to the successful management of leukaemia, a hematological malignancy that disrupts normal blood cell formation. Conventional diagnostic practice, which relies on morphological examination of peripheral blood smears by hematopathologists, is inherently labor-intensive, time-consuming, and subject to inter-observer variability. To address these limitations, this study explores the application of deep learning techniques for the automated classification of malignant lymphoblasts and healthy lymphocytes from microscopic smear images. Four pre-trained Convolutional Neural Network (CNN) architectures-VGG16, ResNet50, InceptionV3, and EfficientNet-B4-were fine-tuned using the publicly available Acute Lymphoblastic Leukaemia Image Database (ALL-IDB). A standardized training methodology was implemented, comprising data augmentation and a two-phase optimization strategy. Comparative analysis revealed that the optimized EfficientNet-B4 architecture achieved state-of-the-art performance, attaining 98.7% accuracy, 99.1% precision, 98.4 % recall, and a 98.75 % F1-score. These findings highlight the potential of optimized CNN models as effective diagnostic support tools, offering enhanced accuracy, reduced diagnostic variability, and promising implications for improving patient outcomes in leukaemia management.

Перевод пока недоступен