Malaria under the microscope: interpretable deep learning for low-resource diagnostic support

Abstract

This work investigates the use of deep learning models for the automated detection of malariainfected blood cells in microscopic images. Two deep learning architectures are evaluated: ResNet50, a widely used convolutional neural network, and Swin Transformer, a more recent model based on hierarchical attention mechanisms. The experimental process includes cross-validation with 5 folds and hyperparameter optimization using the Optuna framework, allowing comparison of the architectures in terms of sensitivity (recall), loss, robustness, and stability between folds. After optimization, Swin Transformer outperformed ResNet50, achieving an average recall close to 0.991. To make the models’ decisions more transparent and aid in clinical interpretation, this study incorporates explainability analysis (XAI) methods, including Grad-CAM, Integrated Gradients, and Occlusion. These methods allow identification of which cell regions contributed to the classification as infected or not. To integrate the entire pipeline - segmentation, classification, explainability, and export of results - the MUM-XAI (Malaria Under Microscope - Explainable AI) application was developed. The system allows for cell-by-cell analysis, visualization of segmentation masks, heat maps generated by XAI methods, and cells highlighted as infected, in addition to generating consolidated reports. Finally, the work discusses the challenge of generalizing the models to a new dataset obtained with a low-cost microscope and without Giemsa staining, highlighting avenues for future research and practical application in low-infrastructure scenarios.