LEUKENet: An Explainable Deep Ensemble Framework for Automated Leukemia Detection Using Segmented and Nonsegmented Microscopy Images

Acute lymphoblastic leukemia (ALL) is a critical hematological malignancy that requires rapid and accurate diagnosis. Most existing automated detection methods rely on segmented datasets and exhibit limited generalization to clinical microscopy images. This study aims to develop and evaluate LEUKENet, an explainable deep ensemble framework for robust, accurate leukemia detection across both segmented and nonsegmented images. It integrates three convolutional neural network architectures—DenseNet‐121, MobileNetV2, and VGG16—into an ensemble classifier. The model was trained on both versions of the ALL‐IDB2 dataset and subsequently tested on two independent datasets (ALL‐IDB1 and blood cell cancer) to assess cross‐dataset generalization. Cell‐level explainability was achieved using Grad‐CAM to visualize discriminative regions within leukemic cells. Statistical robustness was evaluated using 1000 bootstrap resamples. Furthermore, the complete system was deployed as a lightweight web‐based clinical decision support tool. LEUKENet achieved 98.3% accuracy and 99.22% AUC on the segmented ALL‐IDB2 dataset. It generalized effectively to unseen datasets, achieving 93.52% and 90.04% accuracy on nonsegmented ALL‐IDB1 and blood cell cancer images, respectively. Bootstrap analysis confirmed statistically significant improvements ( p < 0.001 against chance; nonoverlapping 95% CIs between configurations) across key performance metrics. Our model demonstrates benchmark performance, strong cross‐dataset generalization, and statistical reliability, along with a practical diagnostic support solution for clinical and resource‐limited healthcare environments.