From Binary Labels to Stage-Aware Bounding Boxes: Evaluating Annotation Depth in Malaria Microscopy

Automated malaria microscopy has progressed from binary parasite classification to-ward object detection and developmental-stage recognition, but the effect of annotation granularity on detection behavior remains insufficiently characterized. We investigate this question using a common YOLOv8m detector under three supervision regimes constructed from the BBBC041 malaria microscopy dataset: binary infection-status detection (R1), parasite-only localization (R2), and five-class stage-aware parasite detection (R3). All regimes are evaluated on the same frozen 120-image test set containing 308 parasite objects. R1 achieves 72.78% precision, 83.53% recall, 77.34% mAP50, and 62.89% mAP50-95. R2 achieves 72.67% precision, 63.31% recall, 68.52% mAP50, and 53.92% mAP50-95. The controlled R2-to-R3 comparison shows the largest performance change: mAP50 decreases by 62.96% and mAP50-95 by 62.33%, while recall decreases by 10.38%. Object-level analysis of R3 shows that 78.57% of ground-truth parasite objects achieve a spatial match at IoU ≥ 0.50, but only 50.32% are both spatially matched and assigned the correct developmental stage. Among spatially matched objects, 35.95% are stage-confused. Bootstrap analysis provides image-level uncertainty intervals, while Grad-CAM visualization qualitatively examines representative correct and stage-confused detections. Frozen cross-dataset tests without external training or threshold tuning show severe transfer degradation on Lacuna and NIH cohorts. The results indicate that finer semantic supervision introduces substantial stage-discrimination difficulty beyond parasite localization alone. Rather than treating annotation richness as intrinsically beneficial or detrimental, our findings motivate evaluating annotation granularity together with class support, label quality, and morphological separability.