Benchmarking learned non-Cartesian k-space trajectories and reconstruction networks

TL;DR

This paper evaluates and compares methods for jointly learning non-Cartesian k-space trajectories and reconstruction networks, highlighting the benefits of projected gradient descent and introducing the HybLearn framework for improved performance.

Contribution

It introduces the HybLearn framework for joint learning and benchmarking of non-Cartesian trajectories and reconstruction, emphasizing hardware constraint enforcement.

Findings

HybLearn outperforms existing methods in joint learning tasks.

Projected gradient descent better enforces hardware constraints than penalty methods.

Benchmarking results on fastMRI dataset demonstrate the effectiveness of proposed approaches.

Abstract

We benchmark the current existing methods to jointly learn non-Cartesian k-space trajectory and reconstruction: PILOT, BJORK, and compare them with those obtained from the recently developed generalized hybrid learning (HybLearn) framework. We present the advantages of using projected gradient descent to enforce MR scanner hardware constraints as compared to using added penalties in the cost function. Further, we use the novel HybLearn scheme to jointly learn and compare our results through a retrospective study on fastMRI validation dataset.