Fig. 1.

Architecture of our proposed approach to scan-specific learning-based SRR. The gray boxes represent the input data, comprising an initial guess of the HR reconstruction, and n acquired LR images. All images and representations are volumetric data in the pipeline. The generative network offers an HR image based on an initial guess. The degradation networks degrade the output of the generative network to fit the LR inputs, respectively, with a mean squared error (MSE) loss. A total variation (TV) criterion is used to regularize the generative network. All losses are combined as a measure for the optimization. Only the generative network is updated during the optimization. The initial guess is obtained from an image reconstructed by a standard TV-based SRR method. The training allows for the SRR tailored to an individual patient as it is conducted on the LR images acquired from a specific patient (no auxiliary datasets of HR images are required). Once the training has been completed, the output of the generative network is taken as the HR reconstruction.