Deep Learning-Based Malaria Parasite Image Classification on Real Microscopy Data
Background: Accurate and timely malaria diagnosis with species-level identification of Plasmodium parasites is critical for guiding effective treatment and disease management. Although light microscopy remains the diagnostic gold standard, its dependence on highly trained personnel limits accessibility, particularly in resource-constrained settings. Recent advances in deep learning have enabled automated image-based diagnosis with promising performance; however, reliable differentiation among Plasmodium species continues to pose a major challenge. This study aims to evaluate and compare the effectiveness of different deep learning architectures for automated species identification from microscopy images. Methods: Three deep learning architectures were systematically assessed: a convolutional backbone (ResNet50), a Vision Transformer (ViT), and a hybrid ResNetViT model. All models were trained from scratch, without using pretrained weights, on a dataset comprising real-world thick blood smear images augmented with publicly available microscopy data from Kaggle. In the hybrid architecture, the ResNet module was used to extract robust local morphological features, while the ViT component captured long-range dependencies and contextual relationships within images. Results: In cross-validation experiments, all three architectures consistently achieved high diagnostic performance. ResNet50 attained the highest accuracy (96.9%), an F1-score of 96.3%, and an ROC-AUC of 0.997. The Vision Transformer achieved 93.1% accuracy, 91.7% F1-score, and an ROC-AUC of 0.989. The hybrid ResNetViT reached 95.2% accuracy, 94.2% F1-score, and an ROC-AUC of 0.995. These results confirm that all architectures can reliably distinguish among Plasmodium species. Although ResNet50 achieved the highest raw accuracy, the hybrid model showed the most stable calibration and the closest qualitative agreement between its attention maps and expert-annotated parasite locations. Conclusions: The findings demonstrate that convolutional, transformer-based, and hybrid deep learning architectures can be successfully trained on real microscopy data for species-level malaria diagnosis. These results support the feasibility of deploying scalable, automated diagnostic systems to improve both accuracy and accessibility of malaria detection, particularly in resource-limited healthcare settings.