Abstract:

The 4D-CT image of the lungs is greatly deformed by breathing and heartbeat, and the motion scale within the lungs may be larger than the structures of interest (blood vessels, airways, etc.) used by the algorithm for optimization. This may lead to the registration algorithm aligning only obvious features such as blood vessels and airways. For the problem of large intensity differences in the registration of lung parenchyma contours, a multi-scale residual deformable image registration framework based on unsupervised end-to-end deep learning is proposed. A multi-scale deep residual network with encoder-decoder structure is used as the deformation field generation model in the proposed registration framework, which enhances feature representation ability, utilizes parameters more efficiently, and effectively improves the convergence ability of the network. The network's ability to perceive multi-scale information is improved through the multi-resolution self-attention fusion module, and a skip connection containing a feature correction extraction module is designed to selectively extract the feature maps output by the encoder and realign them for the decoder to learn alignment offsets. To evaluate the effectiveness of the proposed registration framework, the target registration error of the proposed method on the dir-lab public dataset was compared with traditional methods and current advanced unsupervised registration methods. The results show that the proposed registration framework achieves a target registration error of 1.44±1.24mm on the dir-lab public dataset, which is superior to traditional methods and mainstream unsupervised registration algorithms. In addition, with a controlled folding voxel of less than 0.1%, estimating the dense deformation vector field takes less than 2 seconds, demonstrating the great potential of this algorithm in time-sensitive lung research.

Key words: deep learning, lung CT image, image registration, unsupervised learning