Optimizing U-Net for Brain Tumor Segmentation

Accurate segmentation of brain tumors in magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, and monitoring. Manual annotation is labor-intensive and prone to variability, which motivates the development of automated, robust segmentation techniques. Despite the popularity of U-Net in biomedical imaging, there is limited systematic evaluation of how training strategies and optimization techniques affect its performance in brain tumor segmentation tasks. This study addresses that gap by optimizing the U-Net architecture through hyperparameter tuning and data augmentation, and comparing two training paradigms: training from scratch versus fine-tuning a pre-trained model with a ResNet34 encoder. Experiments were conducted on a publicly available brain tumor MRI dataset. The methodology involved systematic variation of batch sizes, learning rates, and training epochs, combined with augmentation techniques such as rotation, noise injection, and contrast adjustments. The results demonstrate that careful hyperparameter selection significantly improves segmentation accuracy, with the model trained from scratch slightly outperforming the fine-tuned counterpart. However, the fine-tuned model converged faster, suggesting practical advantages in certain scenarios. These findings highlight the importance of tailored optimization in medical image segmentation and support the continued use of U-Net as a strong baseline, especially in data-constrained clinical contexts.