Enhancing Malaria Diagnosis: A Hybrid Deep Learning Approach Integrating CNN-Swin Transformer, and Grad-CAM

Abstract

Malaria remains one of the most pressing global health challenges, demanding diagnostic solutions that are accurate, interpretable, and deployable in clinical practice. Most existing studies rely on traditional machine learning, which depends on handcrafted features and struggles to capture the heterogeneous morphology of malaria parasites. Deep learning approaches, particularly Convolutional Neural Networks (CNNs), have improved feature extraction but are limited by their local receptive fields, weak generalization across diverse blood smear images, and the absence of interpretability mechanisms. Moreover, only limited efforts have been made to utilize red blood cell (RBC) images for robust detection, and prior works seldom explored hybrid architectures or explainable AI (XAI), both of which are crucial for healthcare adoption. To address these gaps, this study proposes a hybrid deep learning framework that integrates CNNs for local feature representation with the Swin Transformer to capture global contextual dependencies. Additionally, XAI techniques, including Grad-CAM, are employed to provide transparent visual explanations by highlighting image regions most relevant to predictions. Experiments conducted on a benchmark malaria cell image dataset demonstrate state-of-the-art performance, achieving 96.8% accuracy, 96.7% precision, 96.9% recall, and a 96.8% F1-score. Confusion matrix analyses confirm balanced sensitivity and specificity, while interpretability visualizations validate the model’s reliability in reducing each false negative and false positive. The results establish the proposed CNN–Swin Transformer with XAI integration as the first effective RBC-based hybrid framework for malaria detection, offering a robust, accurate, and interpretable diagnostic solution well-suited for both advanced clinical workflows and resource-constrained healthcare environments.