A Comparative Analysis of Lightweight Deep Learning Architectures for Malaria Cell Image Classification on Edge Devices

Malaria remains a significant global health burden, particularly in resource-limited regions where automated diagnostic tools are critically needed. This paper presents a comprehensive comparative analysis of lightweight deep learning architectures designed for malaria parasite detection from blood cell microscopy images, with emphasis on deployment on resourceconstrained edge devices. We evaluate five prominent lightweight architectures-MobileNetV3, SqueezeNet, EfficientNetV2, ResNet18, and ShuffleNetV2-across key performance metrics including accuracy, model size, inference latency, and computational complexity. Additionally, we examine model compression techniques such as quantizationaware training and structured pruning, and discuss the implications of deploying these models on edge platforms including Raspberry Pi and NVIDIA Jetson devices. Our findings demonstrate that SqueezeNet variants, particularly a proposed ultra-lightweight variant with 54× parameter reduction, can achieve 94.76% accuracy while maintaining inference speeds suitable for real-time diagnosis in resource-limited settings. Transfer learning provides significant accuracy improvements (3-8% boost) across all architectures. We conclude that the selection of optimal architecture depends on balancing accuracy requirements with computational constraints, with MobileNetV3 and SqueezeNet variants offering the most practical trade-offs for deployment in low-resource healthcare environments[1][4][5][6].