UDRNet: unsupervised deformable registration network of lung CT images with hybrid attention mechanism

With the continuous updates and iterations of diagnostic equipment and technologies, the diagnosis of lung diseases has shifted from single-time-point imaging to multi-time-point imaging data, and from single-modal diagnostic data to multi-modal diagnostic data. However, during this process, factors such as respiratory motion and organ deformation pose challenges for tracking the same lung lesion across multiple time points or modalities, as well as for observing its progression trends. Therefore, to address the challenge of tracking the same lesion region in lung images across different states, we proposes an unsupervised deformable registration network of lung CT images with hybrid attention mechanism. The model directly predicts the deformation vector field (DVF) through an end-to-end encoder-decoder architecture, solving the problems of time consumption and dependence on annotated data in traditional methods. Specifically, we design a Spatial and Channel Hybrid Attention Fusion Module (scHAF) to fuse shallow spatial and channel features in skip connections, enhancing the model's semantic alignment ability and improving the learning of registration-relevant region features. Meanwhile, we design an unsupervised training strategy that optimizes the model using image similarity loss, avoiding the reliance on real deformation field labels. Finally, extensive experiments on the CT Lung Registration dataset demonstrate that our model outperforms baseline methods like 3D VoxelMorph in metrics such as Dice (54.92%), NCC (91.49%), and MSE (89.90%). Further ablation experiments confirm the effectiveness of modules such as scHAF.