Malaria remains a critical global health challenge with over 200 million annual cases, yet traditional microscopic diagnosis is time-consuming and requires expert pathologists. This study proposes a multi-model hybrid framework combining YOLO (v10-v12) for detection and six CNN architectures (DenseNet121, EfficientNet-B0/B1/B2, ResNet50/101) for classification, validated on two public MP-IDB datasets comprising 418 images across 8 classes (4 Plasmodium species and 4 lifecycle stages). The proposed shared classification architecture generates ground truth crops once and reuses them across all detection methods, enabling efficient model training. YOLOv11 achieves 93.10% mAP@50 with 92.26% recall on species detection and 92.90% mAP@50 with 90.37% recall on lifecycle stages. For classification, EfficientNet-B1 reaches 98.80% accuracy (93.18% balanced accuracy) on species despite 45:1 class imbalance, while EfficientNet-B0 achieves 94.31% accuracy (69.21% balanced accuracy) on lifecycle stages with 54:1 imbalance. A notable finding is that smaller EfficientNet models (5.3-7.8M parameters) outperform larger ResNet variants (25.6-44.5M parameters), with EfficientNet-B1 achieving 10.5 percentage points higher balanced accuracy than ResNet101 (93.18% vs 82.73%) despite 5.7× fewer parameters, challenging the "deeper is better" paradigm. The system addresses extreme class imbalance through Focal Loss (α=0.25, γ=2.0), achieving minority class F1-scores of 76.9% on P. ovale with perfect recall and 92.3% on schizont, though trophozoite (51.6%) and gametocyte (57.1%) remain challenging. Implemented on consumer-grade GPU hardware, the framework demonstrates practical feasibility for deployment in resource-constrained settings, though clinical use requires validation on larger datasets and field-collected samples.
Afiliasi berasal dari OpenAlex; metadata OAI-PMH jurnal tidak memuatnya. Penulis tanpa ID OpenAlex dicocokkan berdasarkan nama, sehingga dua orang bisa tergabung atau satu orang bisa terpisah.