Improved Brain Tumor Segmentation Using Modified U-Net based on Particle Swarm Optimization Image Enhancement
Shoffan Saifullah, Rafał Dreżewski
Abstract
This study introduces a robust methodology, a Modified U-Net with Particle-Swarm-Optimization-based Image Enhancement, to address the complexities of brain tumor segmentation. Leveraging PSO-based Image Enhancement's adaptive features, our approach achieves superior performance on a dataset of 3064 Brain MRI images, boasting an accuracy of 99.93%, minimal loss (0.0015), and impressive Dice (0.9699) and Jaccard index (0.9421) values for overall images. The method significantly improves segmentation accuracy, as evidenced by the increase of 9.37 p.p. in Dice and 5.3 p.p in the Jaccard index compared to the U-Net basic approach. Comparative analysis with other methods, including Modified U-Net variants, LinkNet, SegNet, Active Contour, and Fuzzy C-Means, consistently demonstrates outperformance. This method advances medical image analysis by providing precise segmentation and paves the way for future research into optimization and extensions for diverse medical imaging applications.