Motivation: Current MRI mask optimization methods require GT images or lack scan-adaptability. In 3D MRI, the optimal sampling pattern varies by axis, making existing methods challenging to apply effectively. Goal(s): To optimize a scan-specific 3D undersampling pattern using only sparse slices from multiple axes, similar to a scout scan (for calibration), to improve reconstruction quality. Approach: Using only a few slices, like scout scan, generate 2D probability maps, which is interpolated and refined with Fourier weights vectors to produce an optimized 3D probability map to sampling mask. Results: Ours achieved high PSNR, SSIM at an 8x reduction ratio, showing superior preservation of detailed structures. Impact: This study enables faster, high-quality 3D MRI with scan-specific undersampling patterns, potentially reducing scan times in clinical settings. It could improve patient experience, increase MRI accessibility, and enhance outcomes when combined with 3D MR reconstruction network research.
Song et al. (Tue,) studied this question.