Abstract
Malaria continues to be a severe health problem across the globe, especially within resource-limited areas which lack both skilled diagnostic personnel and diagnostic equipment. This study investigates the use of deep learning diagnosis for malaria through ConvNeXt models that incorporate transfer learning techniques with data augmentation methods for better model performance and transferability. A total number of 606276 thin blood smear images served as the final augmented dataset after the initial 27558 images underwent augmentation. The ConvNeXt Tiny model, version V1 Tiny, achieved an accuracy of 95.9%.; however, the upgraded V2 Tiny Remod version exceeded this benchmark, reaching 98.1% accuracy. The accuracy rate measured 61.4% for Swin Tiny, ResNet18 reached 62.6%, and ResNet50 obtained 81.4%. The combination of label smoothing with the AdamW optimiser produced a model which exhibited strong robustness as well as generalisability. The enhanced ConvNeXt V2 Tiny model combined with data augmentation, transfer learning techniques and explainability frameworks demonstrate a practical solution for malaria diagnosis that achieves high accuracy despite limitations of access to large datasets and microscopy expertise, often observed in resource-limited regions. The findings highlight the potential for real-time diagnostic applications in remote healthcare facilities and the viability of ConvNeXt models in enhancing malaria diagnosis globally.