Abstract
Malaria remains a critical global health challenge, requiring accurate and efficient diagnostic tools, particularly in developing countries with limited medical expertise. Detecting malaria parasites from red blood cell (RBC) blood smear images is challenging due to subtle color variations, indistinct demarcation lines, and diverse parasite morphologies. While numerous deep learning models address these issues, their high parameter counts often hinder practical deployment. We propose DANet, a lightweight Dilated Attention Network with approximately 2.3 million parameters, designed for robust malaria parasite detection. DANet employs a novel dilated attention mechanism to capture contextual information and highlight critical features in low-contrast smears, achieving an F1-score of 97.86%, accuracy of 97.95%, and an area under the curve-precision recall (AUC-PR) of 0.98 on the NIH Malaria Dataset, comprising 27,558 images (19,290 training, 2756 validation, 5512 test) from 150 infected and 50 healthy individuals. Compared to state-of-the-art models, including convolutional neural network-transformer hybrids, DANet offers superior efficiency, enabling deployment on edge devices like the Raspberry Pi 4. Its robustness is validated through 5-fold cross-validation and Grad-CAM visualizations, demonstrating enhanced interpretability. DANet provides a practical, high-performance solution for automated malaria diagnosis in resource-constrained settings. The source code of the proposed method is available at: https://github.com/asfakali/DANet .