MRI Segmentation of Brain Tumors based on Multi-scale Feature Fusion and Residual U-Net

Abstract:
Brain tumor segmentation is a vital part of medical image processing, which aims to assist doctors in making accurate diagnoses and treatments. It has significant practical value in the field of clinical brain medicine. MRI is the primary method to diagnose brain tumors, but the fuzzy and adhesion of brain medical images make it difficult to obtain accurate results by manual segmentation. Therefore, we propose a residual network with multi-scale feature fusion based on UNet. This model uses encoder, decoder and skip connection architectures to reduce semantic gaps between networks. We selected the Dice coefficient as a metric, and the higher the Dice coefficient, the better the model’s performance. Compared to other U-Net variant models, the Dice coefficient of our model is improved by between 2% and 4%. The experimental results show that our model is superior to other U-Net variant models in the segmentation task of brain tumors.