Enhancing Leukemia Detection: An Automated Approach Using Deep Learning and Ensemble Techniques

Abstract

As leukemia ranks high among the global causes of death, it's crucial to identify it early to enhance the prognosis for patients. The majority of diagnostic processes used today rely on medical experts inspecting samples manually. This is a laborious process that lacks an automated detection mechanism and takes a lot of time. With a focus on acute lymphoblastic leukemia (ALL), this work suggests an automated diagnostic method that uses Deep Learning (DL)-based ensembles to improve leukemia detection and prediction. We propose to utilize a combination of ten DL techniques (ResNet, ResNeXt, SE-ResNet, Inception V3, VGG, and its variants) and three ensemble techniques (Max voting, Averaging, and Stacking) to constitute the leukemia detection models and observe their performances. The ALL IDB benchmark leukemia dataset was evaluated using these techniques, with performances measured across several metrics namely: classification accuracy, F1 score, precision, recall (sensitivity), Kappa index, and ROC-AUC score. The findings from the experiments demonstrate a notable enhancement in leukemia detection performance when utilizing the proposed techniques. In particular, the proposed Ensemble Max Voting technique surpasses all other stateof-the-art detection models in the literature with an accuracy of 100.0% and an F1 score of 0.997. The main achievement of this study is the identification of the most effective method among various models and techniques for detecting leukemia.