Enhancing Brain Tumor MRI Segmentation Accuracy and Efficiency with Optimized U-Net Architecture

Abstract

This study presents an enhanced approach to brain tumor segmentation using an optimized U-Net architecture, focusing on MRI scans. Our research proposes an automated solution that utilizes U-Net to accurately differentiate between tumorous and non-tumorous tissues, addressing the challenges of manual segmentation such as time consumption, accuracy, and inter-observer variability. Our approach to accurately segmenting brain tumors utilizes the BraTS 2019 dataset and involves preprocessing steps that normalize image data. We employ a modified U-Net model that stands out for its depth and integration of multi-inception modules. Our evaluation metrics, including an IoU score of 0.8252 and a low-test loss of approximately 7.075e-05, highlight the high precision of our model in segmenting brain tumors. However, limitations arise from dataset specificity and potential class imbalance, suggesting future work should focus on enhancing generalizability and addressing computational efficiency. Deep learning has been shown to have significant potential in enhancing diagnostic accuracy and treatment planning in neuro-oncology. This, in turn, opens new opportunities for further developments in automated medical image analysis.

Keywords

• Brain tumor
• Image segmentation
• U-Net architecture
• Image processing