Abstract:

Brain tumor segmentation in magnetic resonance imaging (MRI) is a critical task in medical imaging for diagnosis and treatment planning. In this paper, we employ the U-Net convolutional neural network architecture for automated segmentation of brain tumors on MRI scans. We investigate the effect of hyperparameter tuning and data augmentation on U-Net performance, and compare two training strategies: training a U-Net from scratch versus fine-tuning a pre-trained UNet model. Experiments are conducted on a public brain tumor MRI dataset. The U-Net model is optimized through a series of hyperparameter experiments (varying epochs, batch sizes, learning rates) and enhanced with augmentation techniques to improve robustness given limited data. Results show that careful hyperparameter selection yields a high segmentation accuracy, and that a model trained from scratch slightly outperforms a fine-tuned model. The fine-tuned model with a pre-trained ResNet34 encoder converged faster but required domain-specific training to reach comparable performance. These findings demonstrate the effectiveness of U-Net for brain tumor MRI segmentation and highlight the importance of optimization and augmentation. We conclude with observations on the contributions of an optimized U-Net and discuss future directions such as employing larger datasets, more advanced augmentations, and hybrid modeling approaches.