Advancements in Brain Tumor Segmentation Using Modified UNet Architecture

The ability to segment brain tumors from MRI images is essential for enhancing prognosis, treatment planning, and diagnosis. Manual segmentation is labor-intensive and varies significantly between radiologists, making automated methods essential. This study presents a comprehensive analysis and implementation of various U-Net-based brain tumor segmentation architectures. We explored Vanilla U-Net, U-Net with ResNeXt50 backbone, Feature Pyramid Network (FPN), and an enhanced U-Net model incorporating attention mechanisms and Edge Enhancing Diffusion (EED) filtering. Our models were trained and evaluated using public datasets. The models demonstrated significant improvements in segmentation accuracy, with the best-performing model achieving a Dice Similarity Coefficient (DSC) of 97.642 %, Train Loss of 0.219, Validation Loss of 0.216, Train DICE of 0.808, and Validation DICE of 0.977. These results underscore the potential of advanced U-Net architectures in improving brain tumor detection precision and effectiveness segmentation, providing a reliable tool for aiding radiologists in clinical settinos.