EU-Net: A Neuro-Evolutionary Approach for 3D Brain Tumor Segmentation

This paper introduces a genetic algorithm-based method for optimizing the hyperparameters of a Dilated U-Net architecture aimed at performing brain tumor segmentation from medical imaging. The 3D dilated U-Net employs dilated convolutions to improve feature extraction while maintaining resolution, but its effectiveness is highly correlated upon hyperparameters. Our approach utilizes a GA to explore the hyperparameter space, utilizing a fitness function grounded in segmentation accuracy metrics such as the Dice coefficient and Intersection over Union (IoU), assessed on a benchmark dataset of brain MRI scans. Experimental results reveal that the GA method significantly surpasses traditional manual tuning techniques, achieving enhanced segmentation accuracy and robustness across various tumor types. This study not only underscores the potential of genetic algorithms in optimizing deep learning models but also advances automated brain tumor segmentation.