Item

Abstract

Purpose/Introduction: We present a supervised learning approach to automatically generate MR sequences and corresponding reconstruction from scratch without providing sequence programming rules. New framework is tested in human brain measurement at 3T. Subjects and Methods: The entire scanning and reconstruction process is simulated end-to-end as a fully differentiable concatenated sequence of tensor operations. Each tensor operation from the stack implements: RF events (RFE) (i.e. flip angles and phases), gradient moment events (GME) in x and y, delay times, and a weighting for an ADC, acting on the input model spin system (given in terms of PD, T1 and T2, and DB0). At the sequence learning step, we use Adam[1] optimizer to find the sequence parameters given the loss function specified with respect to data fidelity and SAR cost terms. Task and target We evaluate the proposed method on a simple task: match a target GRE-derived image by optimizing the gradients and flip angles and putting a penalty on SAR. Target sequence details: transient RF- and gradient-spoiled gradient echo readout with linear phase encoding, 24 9 24/48 9 48, TR/TE/FA = 20 ms/3 ms/5. Low FA decreases image blurring. As task sequences we used the same timing pattern and ADC as the target sequence, where for gradients and flip angles we approached four different tasks: Task 1: clone target GME and RFE, optimize only RFE with SAR penalty Task 2: clone target RFE, set GME to 0, optimize GME only Task 3: set RFE and GME to 0, optimize both RFE and GME Task 4: set RFE and GME to 0, optimize both RFE and GME with SAR penalty Results: The result of task 1 displayed in Fig. 1a–e shows that the optimizer prefers to lower flip angles in the outer k-space lines to reduce SAR by 40% keeping a small image error. Figure 1f–j shows how spatial encoding is learned from scratch (Task 2). Figure 2a–c shows that when learning both RFE and GME, too high flip angles are chosen (Task 3). This can be mitigated by putting an additional penalty on SAR. Figure 3c–e shows the full potential of MRI zero: the invention of a complete MRI sequence (Task 4) that is applicable to image acquisition at a real system (3T Siemens, Prisma) in phantoms and in vivo (Fig. 3). Discussion/Conclusion: We have developed a fully automated MRI sequence generator based on the Bloch equation simulations and supervised learning. While we focus on basic image generation herein, having such a tool at hand paves the way to a novel way of generating optimal MR sequence and reconstruction solely governed by the target provided, which could be a certain MR image, but the possibilities for targets are limitless, e.g. quantification, activation, segmentation or contrasts of other image modalities.