Abstract:Objective To eliminate the influence of motion and obtain accurate myocardial T1 mapping, this research proposes a
motion correction algorithm based on self-supervised deep learning. Methods In this algorithm, a convolutional neural network
was employed to estimate motion field and a spatial transform layer was used to obtain the motion corrected image by applying the
motion field on the moving image. The loss function including the image similarity combined with the smoothness of the motion
field was used to optimize the network. Results In this study, cardiac MRI of 47 healthy volunteers were trained and tested and
compared with a traditional registration method. The results showed that the Dice similarity coefficient, myocardial boundary error
and myocardial T1 quantization error were 0.79, 0.925 mm and 59.22 ms, respectively, which were better than the unregistered
images and traditional method. Conclusion The proposed method provides an efficient and accurate tool for clinical diagnosis of
heart disease, which holds a potential for wide clinical application.