Motion-resolved 3D Pulmonary MRI Reconstruction using Sinusoidal Representation Networks.

BACKGROUND: Deep learning reconstruction for free-breathing pulmonary MRI.

OBJECTIVE: To propose a motion-resolved 3D pulmonary MRI reconstruction scheme using the sinusoidal representation network (SIREN).

METHODS: The proposed scheme learns the registration maps using SIREN to register an averaging image to get the final reconstructions. The learning of the network relies only on the undersampled data from the specific subject. The usage of the network for outputting the registration maps enables a memory-efficient algorithm, as outputting registration maps instead of images only requires small networks. The training of the network based on only undersampled data enables an unsupervised learning scheme, which makes the proposed scheme useful in cases in which fully sampled data is not available.

RESULTS: We compare the proposed SIREN-based motion-resolved reconstruction with two state-of-the-art methods for ten datasets. Both visual and quantitative comparison indicates the better performance of the proposed method.

CONCLUSION: In conclusion, the use of SIREN for 3D pulmonary MRI reconstruction allows for the efficient and accurate reconstruction of data that has been undersampled.