Malaria Parasite Detection through CNN, Transformer, and Hybrid CNN–Transformer Models

Abstract

Malaria remains a major global health challenge, particularly in resource limited regions where reliable diagnostic tools are scarce. Traditional thick blood smear microscopy, while accurate, is time-consuming and requires expert interpretation. This study employed Artificial Intelligence as a pre-diagnostic tool to mitigate the impact of malaria by evaluating six deep learning architectures for classifying malaria-infected cell images. The models include two CNN-based architectures, CNN and AlexNet; two hybrid CNN Transformer architectures, ConvNeXt and MaxViT; and two transformer-based architectures, ViT and Swin Transformer. The dataset consisted of 1,883 Giemsa-stained blood smear samples collected from patients at Chittagong Medical College Hospital, Bangladesh, with expert annotations from the Mahidol Oxford Tropical Medicine Research Unit. All models were trained for 100 epochs using the Adam optimizer, a cosine annealing learning rate scheduler, early stopping, and a batch size of 32 samples. Among the tested models, the hybrid CNN Transformer architectures achieved the best performance, with MaxViT yielding the highest accuracy of 100%, followed by ConvNeXt with 96.63%. These findings demonstrate that combining convolutional and transformer mechanisms yields superior performance compared with single-framework models. With their high accuracy, cost-effectiveness, and computational efficiency, hybrid architectures show strong potential as reliable pre-diagnostic tools for malaria detection in resource-constrained settings.