Sub-second whole brain T2mapping via multiband SENSE multiple overlapping-echo detachment imaging and deep learning

Simin Li,Taishan Kang,Jian Wu,Weikun Chen,Qing Lin,Zhigang Wu,Jiazheng Wang,Congbo Cai,Shuhui Cai
DOI: https://doi.org/10.1088/1361-6560/acfb71
2023-10-05
Abstract:Objective. Most quantitative magnetic resonance imaging (qMRI) methods are time-consuming. Multiple overlapping-echo detachment (MOLED) imaging can achieve quantitative parametric mapping of a single slice within around one hundred milliseconds. Nevertheless, imaging the whole brain, which involves multiple slices, still takes a few seconds. To further accelerate qMRI, we introduce multiband SENSE (MB-SENSE) technology to MOLED to realize simultaneous multi-slice T2mapping.Approach.The multiband MOLED (MB-MOLED) pulse sequence was carried out to acquire raw overlapping-echo signals, and deep learning was utilized to reconstruct T2maps. To address the issue of image quality degradation due to a high multiband factor MB, a plug-and-play (PnP) algorithm with prior denoisers (DRUNet) was applied. U-Net was used for T2map reconstruction. Numerical simulations, water phantom experiments and human brain experiments were conducted to validate our proposed approach.Main results.Numerical simulations show that PnP algorithm effectively improved the quality of reconstructed T2maps at low signal-to-noise ratios. Water phantom experiments indicate that MB-MOLED inherited the advantages of MOLED and its results were in good agreement with the results of reference method.In vivoexperiments for MB = 1, 2, 4 without the PnP algorithm, and 4 with PnP algorithm indicate that the use of PnP algorithm improved the quality of reconstructed T2maps at a high MB. For the first time, with MB = 4, T2mapping of the whole brain was achieved within 600 ms.Significance.MOLED and MB-SENSE can be combined effectively. This method enables sub-second T2mapping of the whole brain. The PnP algorithm can improve the quality of reconstructed T2maps. The novel approach shows significant promise in applications necessitating high temporal resolution, such as functional and dynamic qMRI.
What problem does this paper attempt to address?