Enhancing U-Net Performance in Retinopathy Segmentation using Metaheuristic Search Space Optimization

Image segmentation plays a vital role in medical decision-making, significantly improving the efficiency and effectiveness of healthcare services. Consequently, biomedical image segmentation has emerged as a major area of research within computer vision. The progress in deep learning has resulted in the creation of various manually engineered methods, yielding impressive outcomes and establishing new benchmarks in the field. However, a significant difficulty in effectively implementing these algorithms is the selection of suitable hyperparameters. In this research, we tackled this issue by using four meta-heuristic algorithms: Grey Wolf Optimizer (GWO), Particle Swarm Optimizer (PSO), Bayesian Optimizer (BO) and Reinforcement Learning (RL). By employing these techniques, we automated the optimal hyperparameter selection process, thereby improving image segmentation performance. Our experiments were conducted using Diabetic Retinopathy datasets (AMDs), which comprised 908 training samples and 228 testing samples. The comparative analysis demonstrated that the proposed U-Net-PSO method, which has 276,276 trainable parameters, outperformed the traditional U-Net model, which has 1 million trainable parameters. The accuracy levels achieved by our method were 96.4%, compared to just 75.1% for the traditional U-Net, despite having significantly fewer trainable parameters