Processing math: 100%
1960
Adaptive deep image reconstruction using G-SURE
Hemant Kumar Aggarwal1 and Mathews Jacob1
1University of Iowa, Iowa City, IA, United States

We introduce a novel approach to adapt deep learned reconstruction algorithms to a new forward model. The proposed scheme relies on G-SURE loss metric, which accounts for the noise in the measurements. By minimizing the risk of overfitting, this scheme offers improved reconstructions.

Fig. 1: These are the experiments in 2D settings at 6-fold acceleration with a noise of standard deviation of 0.02. Here, we trained the MoDL only with mask M0 as shown in (b). During testing, we utilized the same mask M0 as well as a different mask M1 (d). The reconstruction results in (c) and (e) show that MoDL architecture is robust to small changes in the sampling mask.
Fig. 2: (a) The Cartesian mask M0 used during training. (b) The regridding reconstruction of the test image using M0. (c) MoDL results in good reconstruction since M0 is used here during testing. (d) This is a different mask M1 to test the robustness of MoDL architecture. (e) is the corresponding regridding reconstruction. The reconstruction (f) has lower PSNR than (c), as expected since M1 was not used during training. (h) is the result of model adaptation on (f) using MSE (i) is the reconstructed image using the G-SURE-based model adaption.