Attention-refined U-Net with Skip Connections for Effective Brain Tumor Segmentation from MRI Images

Brain tumor segmentation in MRI scans is a crucial yet challenging task due to the high variability in tumor shape, size, and location. Accurate and efficient segmentation is of paramount importance for timely diagnosis and effective treatment planning in brain tumor patients. In response to these challenges, this paper presents an enhanced U-Net architecture named SC-SE U-Net. Our proposed model integrates Squeeze-and-Excitation channel attention blocks and skip connections similar to residual network architecture, into the traditional U-Net framework. The Squeeze-and-Excitation blocks enhance the model’s capability to focus on more relevant feature maps and suppress less pertinent ones, thereby improving the overall segmentation accuracy. Simultaneously, the skip connections facilitate the flow of gradients during the training process, leading to more stable and faster convergence, and improved performance. We conducted an extensive evaluation of our proposed model on the TCGA-LGG brain tumor dataset. The results show that the SC-SE U-Net outperforms several existing segmentation methods, achieving an Intersection over Union score of 84.22% and a Dice score of 91.43%. The impressive performance of SC-SE U-Net underscores its potential to significantly improve diagnostic efficiency in clinical settings, highlighting the importance of further research in this direction.