Manual microscopic examination of thick blood smears remains the gold standard for malaria diagnosis in resource-limited settings, yet it is labor-intensive, slow, and significantly prone to human error. This study addresses these diagnostic bottlenecks by leveraging the recently released YOLOv12 object detection framework to automate the simultaneous detection of Plasmodium parasites and White Blood Cells (WBCs). Utilizing the challenging, smartphone-captured Nakasi dataset, the research focuses on bridging the gap between high-performance deep learning and the practical constraints of field-based microscopy. The study design follows a robust methodology centered on stratified 5-fold cross-validation for architecture selection, ensuring that the model generalizes well across diverse samples. To handle the significant morphological variability inherent in field-captured images, targeted offline data augmentation techniques were applied. Contrary to the prevailing trend of deploying increasingly massive neural networks, a comprehensive computational analysis revealed that the ultra-lightweight YOLOv12-Nano variant delivers accuracy comparable to its larger counterparts. Most notably, this variant reduces floating-point operations (FLOPs) by over 90%, making it uniquely suited for deployment on low-cost mobile devices. On a strictly isolated 100-image hold-out test set, the optimized pipeline established a new state-of-the-art benchmark, achieving an overall mAP@50 of 0.795 and an mAP@50–95 of 0.463. Error analysis indicates robust class separation between parasites and leukocytes, although persistent challenges remain in distinguishing faint ring-stage parasites from common staining artifacts. Ultimately, this work demonstrates that modern, highly efficient single-stage detectors offer a viable, scalable pathway for deploying real-time, AI-assisted diagnostics on edge hardware. By providing high-fidelity detection with minimal computational requirements, this approach facilitates more accessible and reliable malaria screening in endemic regions.
Ajayi et al. (2026) studied this question.