Lung image registration is widely used in image-guided radiotherapy, but large deformable registration often takes a long time and is prone to noise-induced abnormal deformations, making it difficult to meet clinical accuracy requirements. To address these issues, we propose a lung CT image registration method that combines intensity features around key points and geometric structure features of local key points.

The proposed registration method consists of four processes: feature extraction, feature correlation coefficient calculation, displacement vector field (DVF) generation, and spatial transformation. 1) Feature extraction is implemented using the MIND feature extraction module to extract feature maps from both the fixed image and the moving image, resulting in fixed and moving image feature maps. 2) An adaptive window feature correlation coefficient calculation module is constructed to automatically find a suitable window for calculating the correlation coefficient through model training, thus accurately describing the correspondence between key points and saving training time. 3) In the displacement vector field generation process, two multiscale and dense connection fusion U-Net neural networks are stacked to generate the predicted displacement vector field. First, a multiscale residual convolutional network is used to extract more detailed feature information at multiple scales, which alleviates the problem of gradient vanishing due to increased network depth. Then, dense connection graph convolution is proposed to extract information combining intensity feature correlation coefficients around key points and geometric structure features between local key points. This combined information is used to characterize the correspondence between points in the image pair, alleviating abnormal deformations in certain regions of the lung image registration caused by relying on intensity features alone, and thereby improving the robustness of the lung image registration. 4) The spatial transformation process is implemented using a spatial transformation network (STN), which warps the moving image feature map to the warped image feature map based on the predicted DVF.

The proposed algorithm is evaluated on the DIR-lab, COPDgene, and Creatis datasets. Experimental results show that, in terms of target registration error, the proposed method achieves average target registration errors of 1.21 mm (Table 1), 1.53 mm (Table 2), and 1.00 mm (Table 3) on the DIR-lab, COPDgene, and Creatis datasets, respectively. Compared to the Graphregnet algorithm, the target registration error (TRE) of the proposed algorithm on the DIR-lab, COPDgene, and Creatis datasets is reduced by 18.8%, 13.1%, and 6.5%, respectively, indicating that the proposed algorithm is effective in large deformation image registration. In addition, the proposed algorithm achieves the highest Dice similarity coefficient (DSC) values and the lowest 95% Hausdorff distance (HD95) values among other algorithms on all three datasets, indicating its superior boundary alignment performance. The percentage of negative Jacobian determinants for the proposed algorithm is 0, indicating good topology preservation and no folding phenomenon in the warped image during registration, while other algorithms show a percentage greater than 0 on the DIR-lab and COPDgene datasets, indicating inferior topology preservation capabilities (Table 5). Compared to the intensity difference heat map between the moving and fixed images, the intensity difference heat map between the warped image and the fixed image shows significant reduction in large-scale differences within the lung parenchyma, indicating alignment of tissues and vessels within the lung parenchyma. In addition, the differences in the inferior border of the lung in the coronal and sagittal planes are significantly reduced after registration compared to before, indicating that the lung contour is well aligned (Fig. 12). Comparing the registration results of Graphregnet and the proposed algorithm, both algorithms show good registration performance overall. However, in the anatomical regions marked by the pink boxes, the warped images generated by the proposed algorithm are more similar to the fixed images, and the registration effect on vascular details is significantly better than that of the Graphregnet algorithm (Fig. 13). The registration time of the proposed algorithm is only 0.57 s, demonstrating its good real-time performance.

We present an adaptive windowing algorithm for the registration of lung images with large deformations. The algorithm efficiently computes the correlation of key point features through an adaptive window feature correlation coefficient computation module. In the network encoding stage, a multi-scale residual convolutional network is used to extract multi-scale features, utilizing residual modules to mitigate gradient vanishing caused by network depth. By integrating local geometric structure information of key points and low-dimensional displacement embedding information, a graph convolutional network is introduced for feature learning. Through a dense graph convolutional neural network, sufficient extraction of intensity information around key points and local geometric structure information is achieved. Experimental results show that the proposed algorithm has good real-time performance and high registration accuracy in DIR-lab, COPDgene, and Creatis datasets, and can effectively reduce the probability of deformation anomalies in the local region of large deformable registration.