Enhanced Unsupervised Image Registration via Dense U-Net and Channel Attention

Abstract

In the realm of critical clinical medical image analysis, particularly in surgical navigation and tumor monitoring, the importance of precise image registration cannot be overstated. Acknowledging the need for improved accuracy in current unsupervised image registration methods for single-modal images, this research presents a groundbreaking deep learning-based algorithm. The core innovation of this algorithm lies in its integration of short and long connections, which establish a densely connected architecture within the U-Net framework. This approach significantly enhances feature map interconnectivity, effectively bridging the semantic gaps caused by varying sampling depths within the feature maps. Additionally, the algorithm introduces a channel attention mechanism within the U-shaped network's decoder, which plays a crucial role in reducing image noise and generating smoother deformation fields. This enhancement not only improves the model's sensitivity to finer details but also substantially increases the precision of image registration, a benefit particularly notable when working with single-modal brain MRI datasets. Extensive clinical trials have demonstrated the algorithm's significant contributions to enhancing the accuracy of medical image registration. In conclusion, by harnessing the power of deep learning and innovative algorithmic design, this study tackles critical challenges in medical image registration, providing more precise and reliable support for clinical applications such as surgical navigation and tumor monitoring.