Deep learning-based detection and classification of acute lymphoblastic leukemia with explainable AI techniques

Abstract

Leukemia is identified by an excess of immature white blood cells (WBC) being formed in the bone marrow, leading to cancer. It is divided into two main types: acute, which stems from early cell growth ab- normalities and involves rapid immature cell proliferation, and chronic, which progresses more slowly due to a blockage in the later stages of the cell life cycle. Detecting acute lymphoblastic leukemia (ALL) at an early stage is critical to reducing its associated mortality rate. This study presents an empirical analysis of various pre-trained deep learning models, including VGG16, VGG19, ResNet50, Xception, ResNet152, EfficientNet- B0, NASNetMobile, DenseNet169, DenseNet121, and EfficientNetV2B0, for the detection and classification of ALL. A comprehensive evaluation highlights the effectiveness of deep learning in distinguishing different types of ALL, demonstrating its potential as a reliable diagnostic tool in medical imaging. Additionally, we evaluated the performance of these models using different optimization techniques, including Adadelta, SGD, RMSprop, and Adam, to determine the most effective optimization strategy for improving classifica-tion accuracy. Our results demonstrate that EfficientNet-B0 achieved a classification accuracy of 72 %, while NASNetMobile attained 81 %. Notably, DenseNet121 outperformed these models with an accuracy of 99 %. Furthermore, the remaining seven models VGG16, VGG19, ResNet50, Xception, ResNet152, DenseNet169, and EfficientNetV2B achieved a perfect classification accuracy of 100 %, highlighting their robustness and effectiveness in our experimental setup. To improve the interpretability of the leukemia detection process, explainable AI techniques, including Grad-CAM, Score-CAM, and Grad-CAM++, were integrated to vi-sualize critical regions influencing model predictions. These techniques enhance transparency by providing visual explanations of classification decisions. A detailed comparative analysis was conducted, examining key parameters such as learning rate, optimization algorithms, and the number of training epochs to determine the most effective approach. The study leveraged a publicly available acute lymphoblastic leukemia dataset to ensure comprehensive model evaluation. By offering insights into model performance and interpretability.