Optimized Deep Learning Approaches for Malaria Cell Detection

Abstract

Malaria remains a leading cause of death in sub-Saharan Africa, with Nigeria accounting for approximately 27% of global malaria deaths. Accurate diagnosis is frequently compromised in rural healthcare facilities due to shortages of trained microscopists and limited access to diagnostic equipment. This study presents a comparative evaluation of three convolutional neural network (CNN) architectures, a Baseline CNN trained from scratch, ResNet50, and MobileNetV2 for automated malaria parasite detection in blood smear images. Models were trained and evaluated on the publicly available Malaria Cell Image Dataset (27,558 segmented cell images) using standardized preprocessing, data augmentation, and training protocols. Performance was assessed through diagnostic metrics (accuracy, sensitivity, specificity, precision, F1 score, AUC-ROC) and deployment feasibility metrics (model size, CPU inference latency). MobileNetV2 achieved 93.52% accuracy and 94.87% sensitivity, beating ResNet50 across all diagnostic metrics while requiring 90.5% fewer parameters (2.2M vs. 23.5M) and achieving 3.4× faster CPU inference (22.0 ms vs. 74.3 ms). The Baseline CNN achieved the highest raw accuracy (95.77%) but requires domain-specific training from scratch, limiting practical deployment in low-data settings. These findings establish MobileNetV2 as the optimal architecture for malaria detection under rural Nigerian infrastructure constraints, demonstrating that deployment feasibility must be prioritized alongside diagnostic accuracy in global health AI applications.