Automatic Brain Tumor Segmentation Using Convolutional Neural Networks: U-Net Framework with PSO-Tuned Hyperparameters

Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) data is imperative for precise diagnosis and treatment planning. Manual segmentation, while accurate, is labor-intensive and subject to human error. In this study, we propose an innovative approach leveraging a modified convolutional neural network (CNN) architecture, U-Net, optimized using Particle Swarm Optimization (PSO) to tackle this challenge. Our method achieves significantly improved segmentation accuracy through pre-training hyperparameter tuning, particularly adjusting learning rates and dropout rates with PSO. Compared to existing methods, we observe enhancements of up to 4 p.p. in the Dice Similarity Coefficient (DSC) and 2 p.p. in the Jaccard Index (JI). Using skip connections and dropout layers in CNN-U-Net enables the effective capture of intricate features while mitigating overfitting, resulting in robust segmentation performance. Experimental results showcase the superiority of our approach across different tumor classes, including Meningioma, Glioma, and Pituitary, as well as overall, with maximum DSC and JI values of 94.14% and 89.02%, respectively. Comparative analysis against established techniques underscores the reliability and robustness of our proposed method. By demonstrating the efficacy of deep learning coupled with metaheuristic optimization in medical image segmentation, our study contributes to advancing the field’s understanding and applications. This research lays a foundation for future automated brain tumor segmentation developments, with implications for clinical practice and patient care.