Particle Swarm-Optimized U-Net Framework for Precise Multimodal Brain Tumor Segmentation
Shoffan Saifullah, Rafał Dreżewski
Abstract
Medical image segmentation for brain tumor analysis requires accurate and efficient models due to complex multimodal MRI data and tumor variability. This study presents PSO-UNet, which integrates Particle Swarm Optimization (PSO) with U-Net for dynamic hyperparameter tuning of filters, kernel size, and learning rate. PSO-UNet achieves state-of-the-art segmentation with DSCs of 0.9578 and 0.9523 and IoU scores of 0.9194 and 0.9097 on BraTS 2021 and Figshare datasets. The model uses only 7.8M parameters and executes in 906 seconds, outperforming standard U-Net frameworks. PSO-UNet generalizes well across modalities and tumor types, offering a lightweight and clinically viable solution. This framework offers a novel integration of automated PSO-driven hyperparameter tuning into U-Net, enhancing segmentation performance while reducing computational overhead. Future work will explore hybrid optimizations and further scalability.