Medical Image Registration with Multi-Dilated Convolution and Quintuple Attention

In this study, we propose a registration model using multi-dilated convolution and quintuple attention. The proposed model compensates for the difficulty of capturing global features by using multi-dilated convolution. Quintuple attention is used to capture important features by weighting them in multiple aspects. Our results are compared with those of VoxelMorph and HyperMorph based on convolution neural networks. The registration was done with the spatial transform. Compared to other models, our model outperformed existing models in several metrics.Clinical Relevance-Medical image registration has applications in the clinical field, such as in patient diagnosis and prognosis prediction using medical images. The proposed model can compensate for the diagnostic difficulties that may occur due to internal deformation caused by organ movement in medical images captured at different times by registering the expiratory and inspiratory states of the lung. Therefore, the proposed model with multiple modules is promising in clinical applications.