Enhanced detection of malaria parasites in resource constraint environment using data-efficient image transformer
Segun Adebayo, Oladipo Olarinre Oladosu, Mary Oboh, Halleluyah Oluwatobi Aworinde, Alaba B Ayenigba, Oludamola V. Adeleke, Bukola Oyerinola Atobatele, Tunde Oladipo, Olumide Thomas Adeleke · Intelligence-Based Medicine · 2026
Malaria remains a major public health challenge, particularly in low- and middle-income countries where access to reliable diagnostic infrastructure is limited. Conventional malaria diagnostic methods, including microscopy and rapid diagnostic tests, are often labor-intensive, prone to human error, and may exhibit reduced sensitivity under low parasite density conditions. These limitations highlight the need for automated and scalable diagnostic solutions suitable for resource-constrained environments. This study investigates the use of Data-efficient Image Transformers (DeiT) for automated malaria parasite detection from microscopic blood smear images. Leveraging transfer learning and few-shot training strategies, lightweight transformer models (DeiT-Tiny and DeiT-Small) were fine-tuned using a limited number of labeled samples to simulate real-world data scarcity scenarios. A comprehensive hyperparameter optimization strategy involving multiple optimizers, learning rates, batch sizes, and regularization settings was conducted to identify optimal configurations for deployment in low-resource settings. Model performance was evaluated using clinically relevant diagnostic metrics including accuracy, sensitivity, specificity, F1-score, ROC-AUC, Matthews correlation coefficient (MCC), and diagnostic odds ratio (DOR). Experimental results demonstrate that the modified DeiT model significantly outperforms baseline transformer architectures, achieving a classification accuracy of 76.7%, sensitivity of 76.19%, specificity of 84.99%, and an ROC-AUC of 0.9006. The model also achieved a diagnostic odds ratio of 18.13, indicating strong diagnostic capability. Despite maintaining a lightweight architecture with only 5.72 million parameters, the proposed model demonstrates improved feature discrimination and classification performance compared to conventional Vision Transformer baselines. Our findings suggest that data-efficient transformer architectures can provide a practical and scalable approach for AI-assisted malaria diagnosis in resource-limited environments. The proposed framework has the potential to support early malaria detection, reduce diagnostic workload, and improve healthcare delivery in malaria-endemic regions.