Automated leukemia detection from microscopic images using deep transfer learning with explainable AI-based analysis.

Leukemia, a life-threatening hematological malignancy, necessitates early and accurate diagnosis for effective treatment. Traditional microscopic examination is often time-consuming and subjective, while most existing deep learning-based studies are limited to a narrow selection of models and lack interpretability analysis. To address these gaps, this study presents a comprehensive comparative evaluation of fourteen pre-trained CNN architectures, including EfficientNetB1/B3/B5, Xception, InceptionV3, InceptionResNetV2, DenseNet201, ResNet variants, VGG16/19, and MobileNet, for automated leukemia detection from microscopic blood smear images. A publicly available Kaggle dataset containing 10,700 labeled images with variations in staining intensity, cell morphology, and image resolution was used. The dataset was divided into training (70%, 7500 images), validation (15%, 1600 images), and testing (15%, 1600 images) subsets using stratified sampling to preserve class balance. To ensure statistical robustness, all experiments were repeated ten times with different random seeds, and performance metrics are reported as mean values, with pairwise McNemar's tests applied to validate the significance of observed differences (p < 0.05). EfficientNetB1, EfficientNetB5, Xception, and ResNet50 achieved test accuracies of up to 96%, recall up to 98%, F1-scores up to 97%, and AUC values up to 0.987, indicating excellent threshold-independent discriminative capability. Explainable AI techniques, specifically Grad-CAM and LIME, were integrated to enhance transparency and clinical trustworthiness. Comparative analysis demonstrates competitive or superior performance against recent state-of-the-art methods, providing preliminary evidence toward future clinical translation. However, the study is limited by reliance on a single publicly available binary-class dataset, which constrains generalizability across imaging devices, staining protocols, and clinical sites; external validation on independent multi-center datasets is therefore identified as a primary direction for future work.