Comparative Analysis of Parallel Particle Swarm Optimization in Lightweight Deep Learning: A MobileNet Case Study for Skin Cancer Detection

Skin Cancer is one of the most common cancers around the world. Detecting it in early stages is essential to improve the patient’s chance of survival. This study introduces a framework designed to improve the classification performance of binary skin cancer classification by tuning the hyperparameters of the MobileNet model using Particle Swarm Optimization (PSO). The optimizer is implemented using two approaches: a traditional sequential PSO and a parallel multi-threaded PSO approach. The ISIC 2018 dataset is used, and the hyperparameters such as the batch size, optimizer selection, and dense layer size were considered for tuning. The findings show that both approaches have achieved a strong classification accuracy between 95% and 97%. The parallel approach, however, has shown 20x up to 46x reduction in the training time, yet this was accompanied by an increase in memory usage, which rose by approximately 20% to 150% compared to the sequential approach, depending on swarm size and iteration count. These findings show the resource trade-off between training time vs memory consumption for applying an optimized Deep Learning Model in clinical settings, especially where computing resources are limited.