부산대학교 도서관

MRI suffers from an inherent trade-off between signal-to-noise ratio, acquisition time, and spatial resolution [1]–[3]. K-space undersampling allows for a reduction in acquisition time at the cost of decreased signal-to-noise ratio and/or spatial resolution [7]. Initially developed to aid in tissue segmentation [18], [19], U-Nets have also been utilized effectively to improve the image quality of a diverse array of MRI contrasts [20]–[22]. The performance of U-Nets can be enhanced by providing the network with a fully sampled complementary contrast as a prior to allow for cross-contrast feature transfer [29]. We implement aim to verify that cross-contrast feature transfer improves the quality of images output by a dense U-Net (DU-Net). We assess whether the choice of complementary T1 or T2 weighted MRI contrasts for undersampling affects the quality of output images. We also aim to improve the sensitivity of image quality metrics used to compare networks by restricting their calculation to areas bounded around regions of clinical diagnostic interest. The quality of DU-Net outputs does not change significantly when the contrast used as a prior is exchanged; image quality metrics are all within one standard deviation of each other. There is also no quantifiable difference between models trained with a cross-contrast prior and those that are not. There are, however, qualitative improvements, particularly in the regions around tumors. Bounding error calculation regions leads to an increase in the significance of the measured difference between DU-Net outputs in most cases.