Deep Learning-based Malaria Parasite Image Classification on Real Microscopy Data

Background and Objective: Accurate and timely malaria diagnosis, including species-level identification of Plasmodium , is essential for guiding effective treatment and disease management. Traditional light microscopy remains the diagnostic gold standard but relies on highly trained personnel, limiting its accessibility in resource-constrained regions. Recent advances in deep learning have enabled automated image-based diagnosis with high performance; however, accurate differentiation among Plasmodium species remains a major challenge. This study aims to evaluate and compare the effectiveness of different deep learning architectures for automated malaria species identification from microscopy images. { Methods : Three architectures were systematically assessed: a convolutional backbone (ResNet-50), a Vision Transformer (ViT), and a hybrid ResNet–ViT framework. All models were trained from scratch, without using pre-trained weights, on a dataset comprising real-world thick blood smear images augmented with publicly available microscopy data from Kaggle. The ResNet module was employed to extract robust local morphological features, while the ViT component captured long-range dependencies and contextual relationships within the images. Results: Across cross-validation experiments, all three architectures demonstrated consistently high diagnostic performance. ResNet‑50 obtained an accuracy of 95.7%, F1‑score 95.2%, and ROC‑AUC 0.997. The ViT model reached an accuracy of 92.9\%, F1‑score 92.6%, and ROC‑AUC 0.992. The hybrid ResNet–ViT achieved an accuracy of 95.2%, F1‑score 94.7%, and ROC‑AUC 0.997. These results confirm that all architectures can reliably distinguish among Plasmodium species, with the hybrid model effectively integrating local and global feature representations. Conclusions: The findings highlight 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 implementing scalable, automated diagnostic systems to enhance accuracy and accessibility of malaria detection, particularly in resource-limited healthcare settings.