Analytical Evaluation of Deep Learning Models for Malaria Parasite Detection in Blood Smear Images from Nigerian Hospitals

Malaria diagnosis in Nigerian hospitals depends heavily on microscopy-based blood smear analysis, a process limited by expert scarcity, interobserver variability, and throughput constraints in high-burden settings. This study analytically evaluates deep learning models for automated malaria parasite detection and quantification in Giemsa-stained blood smear images from Nigerian hospitals. A dataset of 12,400 annotated blood smear images was compiled from three hospitals in Benue State and two in Rivers State, covering P. falciparum and P. vivax species. Three convolutional neural network architectures (ResNet-50, EfficientNet-B4, and a custom lightweight CNN) were trained and compared using transfer learning from ImageNet pre-trained weights. Training used an 80-15-5 train-validation-test split with extensive augmentation to address class imbalance. Evaluation metrics included sensitivity, specificity, and AUC-ROC. EfficientNet-B4 achieved the highest sensitivity of 94.8 percent and specificity of 96.2 percent, with AUC-ROC of 0.982. The lightweight custom CNN achieved competitive performance (sensitivity 91.3 percent, AUC-ROC 0.961) with 40 percent fewer parameters, making it deployable on mid-range smartphones. Error analysis identified dense thick-film samples as the primary source of misclassification. The study fills a research gap in deep learning malaria diagnostics validated on Nigerian clinical data, and proposes integration of the lightweight model into a mobile diagnostic aid for peripheral health facilities with limited lab infrastructure.

Keywords: malaria detection, deep learning, blood smear images, convolutional neural networks, Nigeria