Leuk-XAI-EDL: explainable ensemble deep learning model for leukemia classification
Syed, Saad Ahmed Nisar, Humaira Lee, Yu Jen Hum, Yan Chai Jaffari, Rabeea
Syed, Saad Ahmed
Nisar, Humaira
Lee, Yu Jen
Hum, Yan Chai
Jaffari, Rabeea
Type
Supervisor
Date
2025-11-01
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Leukemia is among the leading causes of death globally, and early detection is critical. Current manual diagnostic methods, reliant on expert examination, are time-consuming and costly. They are also error-prone, particularly with less experienced practitioners. This study contributes to addressing these challenges by presenting an automated approach to detect leukemia using microscopic blood images. The method combines image processing and deep learning (DL)-based ensemble techniques to enhance classification accuracy. While ensemble of DL models can render accuracy, they require significant computational resources that necessitate supercomputing for low latency realizations. This study addresses these challenges by presenting a scalable and explainable ensemble of deep learning framework, favorable for the high-performance computing environments. The method’s reliance on supercomputing is driven by the need to train and evaluate ensemble of robust DL models in parallel, a task that is computationally prohibitive on standard systems due to extensive floating-point operations and memory requirements. First the performance of ten DL baseline models is evaluated. Onward, ensembles of the three best performing models (VGG16, VGG19, and DenseNet121) are designed respectively using max voting, averaging, and stacking. A rigorous performance evaluation on the ALL-IDB dataset confirms that the stacking-based ensemble with a support vector machine meta-learner outperforms the state-of-the-art counterparts by achieving an accuracy of 99.70% and a F1 score of 99.69%. It is demonstrated that leveraging supercomputing capabilities is crucial for managing the parallel workflows and high computational load of ensemble models, enabling the development of accurate, reliable, and clinically viable automated diagnostic tools that can operate at the scale required by contemporary healthcare systems. The interpretability of the findings is enhanced using gradient-weighted class activation mapping and Shapley additive eXplanations techniques.