A deformable patch-based transformer for 3D medical image registration

Medical image registration is of great importance in clinical medicine. However, medical image registration algorithms are still in the development stage due to the challenges posed by the related complex physiological structures. The objective of this study was to design a 3D medical image registration algorithm that satisfies the need for high accuracy and speed of complex physiological structures.

We present a new unsupervised learning algorithm, "DIT-IVNet," for 3D medical image registration. Unlike the more popular convolution-based U-shaped registration network architectures like VoxelMorph, DIT-IVNet uses a combined convolution and transformer network architecture. To better extract image information features and reduce the heavy training parameters, we improved the 2D_Depatch module to a 3D_Depatch module, thus replacing the patch embedding in the original Vision Transformer which adaptively performs patch embedding based on 3D image structure information. We also designed inception blocks in the down-sampling part of the network to help coordinate feature learning from images to different scales.

Dice score, Negative Jacobian determinant, Hausdorff distance, and Structural Similarity evaluation metrics were used to evaluate the registration effects. The results showed that our proposed network had the best metric results compared with some state-of-the-art methods. Moreover, our network obtained the highest Dice score in the generalization experiments which indicated better generalizability of our model.

We proposed an unsupervised registration network and evaluated its performance in deformable medical image registration. The results of the evaluation metrics showed that the network structure outperformed state-of-the-art methods for the registration of brain datasets.