Performance Comparison of GPipe-Optimized VGG16 and LeNet Networks for Malaria Microscopy Image Classification

Abstract

This study addresses the memory–accuracy trade-off that limits deep-learning malaria diagnostics on commodity hardware. This paper curated 27 558 high-quality thin-smear images from the public NLM corpus, split 80/20 for training-test, and benchmarked two convolutional architectures. A compact LeNet was trained on one RTX 2080 Ti, whereas a VGG16 was trained in single-GPU mode and with GPipe pipeline parallelism across two, three and four identical GPUs. The evaluation used accuracy, F1-score, ROC-AUC, throughput and peak memory. LeNet converged in 10 epochs to 77 % accuracy and 0.78 F1 while consuming only 1.9 GB, but its capacity proved insufficient for reliable diagnosis. VGG16 reached 96 % accuracy, 0.96 F1 and 0.99 AUC yet required 22.7 GB on a single card. GPipe redistributed the model, cutting per-GPU memory to about 3 GB with negligible accuracy loss; communication overhead, however, limited throughput scaling. These findings confirm that pipeline model parallelism enables state-of-the-art performance on affordable multi-GPU rigs and outline optimizations—such as finer stage granularity and adaptive batch sizing—to further accelerate deployment in low-resource laboratories. The proposed workflow offers a transferable template for other medium-scale medical imaging tasks.