<![CDATA[Malaria remains a major global health threat and requires rapid, accurate diagnosis. Manual microscopic examination is time-consuming and susceptible to human error, creating a need for reliable artificial-intelligence-based automation. This study developed a computationally lightweight custom convolutional neural network (CNN) for classifying red blood cell images without relying on large transfer-learning architectures and integrated explainable artificial intelligence to improve the transparency of its predictions. The experiments used the balanced National Institutes of Health malaria dataset containing 27,558 images, equally divided into parasitized and uninfected classes. Images were resized to 128 x 128 pixels, normalized, augmented, and used to train basic and optimized custom CNN variants. The optimized architecture incorporated a dropout layer of 0.5, the Adam optimizer with a learning rate of 0.001, binary cross-entropy loss, and early stopping. A clinical decision threshold of 0.8 was applied to reduce false-negative predictions for parasitized cells. The best model was obtained at epoch 6, with a validation accuracy of 94.07% and a validation loss of 0.1659. At the clinical threshold, the model achieved an overall accuracy of 94%, a recall of 98% for the parasitized class, and an F1-score of 94%, with only 59 false negatives among 2,755 parasitized test images. Grad-CAM and saliency-map visualizations showed that the network focused on parasite-related regions within infected cells, providing evidence that the predictions were based on clinically relevant visual features. The results demonstrate that a purpose-built custom CNN can provide accurate, computationally efficient, and interpretable malaria detection, with potential for deployment in resource-constrained and remote healthcare settings.]]>
Copyrights © 2026
