LeukoXAI-Lite: A flexible and modular software framework for the interpretable diagnosis of Acute Lymphoblastic Leukemia by deep learning algorithms and visual explanation tools. The system adopts EfficientNetB3-based convolutional neural network, which is embedded in a hierarchical federated learning framework. This approach facilitates distributed model training with simulated health participants cooperating yet guarantee the private of sensitive patient information. In addition to disease categorization, our framework is equipped with a profound explainable artificial intelligence module, based upon 18 distinct visualization methods that includes saliency maps, guided backpropagation, gradientbased methods, SmoothGrad, VarGrad, SquareGrad, Grad-CAM, Grad-CAM++, HiResCAM, Respond-CAM, Score-CAM, Faster Score-CAM, oclusion sensitivity, LIME, SHAP, sobol attribution, and a fusion approach. These approaches produce visual heatmaps which highlight diagnostically important regions in microscopic images of blood cells, making the model more interpretable for clinical deployment. LeukoXAI-Lite also comes with instruments for systematic evaluation of the performance of the predictive model and explanation methods. We support common classification on based metrics (accuracy, precision, recall, F1 score, Kappa score and MCC) as well the explanation specific ones like Deletion, Insertion, Fidelity and Stability. It is implemented with open-source python libraries to be lightweight, adaptable and compatible with real-world use in medical imaging. LeukoXAI-Lite facilitates such kind of trustworthy and interpretable artificial intelligence solutions for the clinical diagnostics by means promoting transparency, reproducibility and privacy friendly learning. • Secure & Explainable AI for ALL Diagnosis: LeukoXAI-Lite integrates deep learning with Hierarchical Federated Learning (HFL) for privacy-preserving, decentralized Acute Lymphoblastic Leukemia (ALL) diagnosis. • Comprehensive XAI & Visual Interpretation: Features a robust Explainable AI (XAI) module with multiple saliency methods generating interpretable heatmaps for diagnostic significance. • Reproducible & Benchmarked Framework: The open-source Python framework supports rigorous performance benchmarking of both classification and XAI modules using standard and XAI-specific metrics, ensuring transparency and research facilitation.
Parwez et al. (Wed,) studied this question.