可变形图像配准, 卷积神经网络, 注意力机制, 金字塔结构

To address the prevalent issue of insufficient accuracy in current medical image registration algorithms, a pyramid-structured dual-stream decoder-assisted registration algorithm is designed. This algorithm combines the local dependency characteristics of convolutional neural networks with the global dependency modeling capability of the attention mechanism. Through its unique dual-stream decoder design, it achieves progressive fine registration of magnetic resonance brain images. Unlike traditional methods that simply concatenate the images to be registered and then process them, this registration algorithm cleverly combines the advantages of cross-attention calculation and channel dimension concatenation for feature fusion. It can identify various deformation patterns and select the appropriate deformation field. By employing a pyramid structure and neighborhood attention mechanism, it greatly reduces the computational load while ensuring performance. To verify the effectiveness of the algorithm, comprehensive experiments are conducted on two 3D brain MRI (Magnetic Resonance Imaging ) datasets, LPBA ( LONI Probabilistic Brain Atlas ) and Mindboggle. The experimental results show that compared to commonly used registration algorithms, this method has achieved state-of-the-art performance on multiple evaluation metrics, fully demonstrating the strong capability and application potential of the model in deformable medical image registration.

deformable image registration, convolutional neural network, attention mechanism, pyramid structure