Improving U-Net Performance for Tumor Segmentation Using Attention Mechanisms

Abstract

U-Net is a widely recognized neural network model for medical image segmentation, renowned for its efficiency in extracting features from both current and past input data. However, traditional U-Net models exhibit limitations in extracting edge features, particularly in medical CT images characterized by complex gray distributions and close pixel intervals. This leads to suboptimal performance, with low accuracy, recall, intersection over union (IoU), and F1-score. This research proposes an improved U-Net model incorporating an attention mechanism to enhance tumor segmentation accuracy and efficiency. The attention mechanism strategically weights important features, directing the network to focus on task-relevant areas. Experimental results demonstrate that our proposed attention-based U-Net model significantly improves tumor segmentation performance, achieving notable enhancements in accuracy, recall, IoU, and F1-score. Further validation across diverse datasets confirms the model's generalization ability and superiority compared to the original U-Net method. This research contributes to the advancement of medical image segmentation techniques, highlighting the potential of attention mechanisms in optimizing deep learning models for clinical applications.