Abstract:
Accurate and efficient detection of malaria parasites in stained blood smear images remains a critical challenge, particularly
in resource-limited settings where expert microscopists may be unavailable. This study compares two deep learning instance
segmentation models, YOLOv8 and Mask R-CNN, for automated detection, segmentation, and life-stage classification of malaria parasites in publicly available Giemsa-stained microscopy images. A total of 1,328 annotated images were used
to fine-tune YOLOv8n and Mask R-CNN (ResNet-50-FPN backbone). YOLOv8 achieved higher detection performance
with bounding-box mAP50 of 0.648, mask mAP50 of 0.624, mean accuracy of 96.7%, and F1-score of 0.71, compared
to Mask R-CNN’s mAP50 of 0.511, accuracy of 93.2%, and F1-score of 0.48. Bootstrap resampling (1,000 iterations)
confirmed the statistical reliability of performance differences with 95% confidence intervals. YOLOv8 also achieved
faster inference (9 ms per image) than Mask R-CNN (93 ms), highlighting its potential for real-time screening. Despite
data imbalance among parasite stages, both models produced meaningful segmentation masks enabling quantitative
morphological analysis. These results demonstrate that lightweight, statistically validated deep learning architectures
can deliver reliable, scalable, and interpretable tools for automated malaria detection and quantification, promoting AI
integration into diagnostic microscopy workflows.
