多尺度低秩加稀疏模型在加速超高场脑部4D Flow成像中的应用

Multiscale low-rank plus sparsity modeling in fast ultra-high-field cerebrovascular 4D Flow imaging

目的探讨多尺度低秩加稀疏模型在加速超高场脑部4D Flow成像中的应用。方法前瞻性收集2022年10月至2023年1月在复旦大学附属华山医院招募的健康志愿者10名,男5名、女5名,年龄23~35(29±4)岁。基于多尺度低秩(MLR)模型提出针对4D Flow数据特点的多尺度低秩加稀疏(MLRS)模型加速采集和重建算法。首先采用7.0 T MR对健康志愿者进行全采样脑部4D Flow扫描,并对采集的数据进行不同加速倍数(R分别为4、8、12、16)的高斯分布欠采样。计算不同加速倍数下压缩感知(CS)模型、低秩加稀疏(L+S)模型、MLR模型和MLRS模型重建结果与全采样参考图像在血管掩模内的均方误差(RMSE)和峰值信噪比(PSNR),模型间的比较采用配对样本t检验或Wilcoxon带符号秩检验。采用Pearson检验评估不同加速倍数下4种模型的血流动力学参数与全采样参考值的相关性,相关系数的比较采用Wilcoxon带符号秩检验。结果相同加速倍数下RMSE从小到大依次为MLRS、MLR、L+S、CS模型,且MLRS模型的RMSE显著低于MLR、L+S、CS模型(P<0.05);PSNR由大到小依次为MLRS、MLR、L+S、CS模型,且MLRS模型的图像PSNR显著高于MLR、L+S、CS模型(P<0.05)。不同加速倍数下MLRS模型测得的脑血管流速与全采样参考值相关系数显著高于MLR、L+S、CS模型与全采样参考值相关系数(P<0.05)。结论该文提出的MLRS模型在保证图像质量的前提下能够加速超高场脑部4D Flow MR成像;并且相同加速倍数下与传统加速模型相比,MLRS模型具有更高的重建准确度。

Objective To investigate the application of multiscale low-rank plus sparsity(MLRS)modeling in fast ultra-high-field intracranial 4D Flow imaging.Methods Ten healthy volunteers,5 males and 5 females,aged 23-35(29±4)years old,recruited from October 2022 to January 2023 at Huashan Hospital of Fudan University,were prospectively collected.A MLRS model acceleration algorithm was proposed according to the characteristics of 4D Flow data based on the multiscale low-rank(MLR)model.Firstly,full sampling brain 4D Flow scans were performed on healthy volunteers using 7.0 T MR,and the acquired data were under-sampled with Gaussian distributions at different acceleration rates(R of 4,8,12,and 16,respectively).The root mean square error(RMSE)and peak signal-to-noise ratio(PSNR)of the compressed sensing algorithm(CS),low-rank plus sparse algorithm(L+S),MLR,and MLRS model were calculated at different acceleration rates,with fully sampled data as reference.And the comparison of models was performed using the paired-samples t-test or Wilcoxon signed rank test.Pearson′s test was used to assess the correlation between hemodynamic parameters of the 4 algorithms and the fully sampled reference values at different acceleration rates,and the correlation coefficients were compared using Wilcoxon signed rank test.Results The RMSE under the same acceleration rates was MLRS,MLR,L+S,and CS models in ascending order,and the RMSE of the MLRS model was significantly lower than that of the MLR,L+S,and CS models(P<0.05);the PSNR was MLRS,MLR,L+S,and CS models in descending order,and the PSNR of the MLRS model was significantly higher than that of the MLR,L+S,and CS model(P<0.05).The correlation coefficients between the blood flow velocity measured by the MLRS model and the reference value were significantly higher than those of the MLR,L+S,and CS models for different acceleration rates(P<0.05).Conclusion The proposed MLRS algorithm is capable of accelerating ultra-high-field 4D Flow MR imaging of the brain while guaranteeing the image quality,and the MLRS model has higher reconstruction accuracy compared with conventional acceleration models at the same acceleration rate.