Improving MRI series correspondence using a modified TransMorph neural network

Improving the quality of the training sample is an important aspect when training artificial intelligence modules to solve medical problems. In this study, we applied the TransMorph neural network to solve the DIR problem with respect to MRI series. We modified TransMorph using an additional convolutional layer with settings similar to the Sobel filter. The modification made it possible to improve the result of solving the DIR problem relative to the original TransMorph architecture. The results of comparing the original and modified models based on their application to MRI data of the head and MRI of the male pelvis are presented.