Abstract
Malaria remains a major global health concern, particularly in regions with limited healthcare infrastructure. Traditional diagnostic methods such as microscopy, rapid diagnostic tests (RDTs), and polymerase chain reaction (PCR) suffer from scalability, sensitivity, and expertise-related limitations, underscoring the need for automated diagnostic strategies. This study investigates deep learning models for malaria detection from blood smear images. Four convolutional neural networks (CNNs), MobileNetV2, VGG19, InceptionV3, and ResNet18, were empirically evaluated, with ResNet18 achieving the highest F1-score of 96.33%. Building on these results, two advanced hybrid architectures, Xception and Inception-ResNetV2, were fine-tuned on a dataset of 27,090 images from the Kaggle malaria collection, attaining classification accuracies of approximately 98% on validation and test sets. Model robustness was further confirmed using an independent dataset from the Harvard Dataverse containing thick smear images captured under varied staining and imaging conditions, where accuracy remained consistently high (97-98%). To enhance interpretability and clinical trust, three explainable artificial intelligence (XAI) techniques, Gradient-weighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-agnostic Explanations (LIME), and SHapley Additive exPlanations (SHAP), were employed. These complementary methods provide spatial, superpixel, and pixel-level transparency into the models' decision-making. Furthermore, representative misdiagnosed samples are presented, wherein these visualization techniques reveal morphological and staining artifacts that led to erroneous predictions, clarifying model failure modes and improving transparency. The proposed AI-based diagnostic framework thus demonstrates high accuracy, interpretability, and generalization, representing a scalable solution for malaria detection in resource-limited healthcare settings.