Optimizing U-Net Architecture for Brain Tumor Segmentation in MRI: A Comparative Study of Training Parameter Variations

Brain tumor segmentation focuses on differentiating between healthy and tumorous tissues. Early and accurate diagnosis significantly improves the survival rate for individuals affected by this condition. However, manual segmentation of brain tumors in three-dimensional Magnetic Resonance Imaging (MRI) volumes is a labor-intensive and time-consuming process, with accuracy heavily reliant on the operator's expertise. Therefore, there is a strong demand for an accurate and fully automated method to segment brain tumors and measure tumor size. The development and enhancement of Computer-Aided Diagnosis (CAD) systems for this purpose can greatly assist specialists. In this project, we leveraged the power of deep learning networks to address tumor segmentation in brain MRI images. Specifically, we employed the U-Net architecture, which consists of both an encoder and a decoder. This project explores how various training parameters influence the network's segmentation accuracy in a two-dimensional context. We conducted six distinct experiments, each with different parameter settings, and compared their results to evaluate performance.

Keywords:

U-Net network , Magnetic Resonance Images , brain tumor , BraTs ۲۰۱۸