Dynamic MRI using SmooThness Regularization on Manifolds (SToRM)

We investigate the problem of efficiently acquiring and reconstructing a dynamic MR image series. We provide a short description of our work here. For further details, please refer to our paper.

Software

The code for l2-SToRM is available here and the data used for the code is shared here.

Outline

Aim: We propose a method to efficiently reconstruct a dynamic MR image series from highly-undersampled measurements in the free-breathing and ungated setting.

Method: We assume that the image frames lie on a low dimensional manifold. We propose a novel navigator-based acquisition scheme which aids the detection of the neighbours of each frame on the manifold. We recover the image series by enforcing similarity between these neighbouring image frames.

Result: Our reconstructions on an in-vivo free-breathing ungated cardiac dataset are of comparable quality to a breath-held gated reconstruction. Both these acquisitions were of similar time duration.

Detailed Description

The inherent slow nature of MR imaging gives rise to highly under-sampled Fourier domain (k-space) measurements, especially in the case of dynamic imaging where a good temporal resolution is desired. Thus, the image recovery is highly ill-posed and the problem is further complicated by the presence of breathing and cardiac motion.

The clinical practice in cardiac MRI is to ask the patient to hold his/her breath and acquire data in the ECG-gated mode. This is a problem for many patients with limited pulomonary capacity and paediatric patients. Moreover, the ECG gating is inefficient in case the patient suffers from arrhythmia. Hence, we propose a method to acquire data in the free-breathing ungated mode and reconstruct it. 

We note that the frames of the image series only depend on a few physiological parameters like the cardiac and respiratory phase. Thus, the images lie on a low-dimensional manifold embedded in high dimensional space. We recover the images by enforcing similarity between neighbouring images on the manifold. In order to identify the neighbours of each frame on the manifold, we propose a novel acquisition scheme containing navigator lines. Our acquisition scheme and reconstruction algorithm are summarized in Fig 1. 

We acquired and reconstructed in-vivo free-breathing ungated cardiac data using our proposed scheme. We also acquired a breath-held gated dataset on the same subject. Our free-breathing reconstructions are seen to be of comparable quality to the breath-held reconstructions. Moreover, the scan times for both acquisitions were the same (~ 3.3 mins for 5 slices). The results are shown in Fig 2. Note that the slice positions from the 2 datasets do not match exactly due to respiratory and cardiac motion.

Reference

Sunrita Poddar and Mathews Jacob. "Dynamic MRI using smooThness regularization on manifolds (SToRM)." IEEE Transactions on Medical Imaging 35.4 (2016): 1106-1115.