Leveraging Pre-Trained CNNs and Ensemble Methods for Improved Diagnosis of Acute Lymphoblastic Leukemia

Acute lymphoblastic leukemia (ALL) is a critical form of blood cancer, with early and accurate diagnosis essential for effective treatment. This study proposes an innovative approach to automate ALL detection by leveraging deep Convolutional Neural Networks (CNNs) and a weighted ensemble learning method. An ensemble of pre-trained CNN architectures—AlexNet, VGG-16, ResNet-50, ResNet-152, MobileNet, and EfficientNet-V2—was used to enhance classification accuracy through a weighted voting mechanism based on performance metrics such as F1- score and area under the curve (AUC). Rigorous preprocessing was applied, including data augmentation and class balancing, using the CNMC-2019 ALL dataset. The model achieved a high weighted F1-score of 97.01% and a balanced accuracy of 97.54%, demonstrating superior performance in distinguishing leukemic cells. Furthermore, gradient class activation mapping was employed for model interpretability, highlighting relevant image areas. The approach shows significant promise for early ALL detection, enhancing clinical decision-making and ultimately improving patient outcomes.