To accelerate MRF acquisition, in this study a sliding-window reconstruction strategy is proposed to improve the quality of each image in MRF by reducing the undersampling artifacts, thus decreasing the necessary number of time points (L) ¹.

For the original MRF method², each spiral interleaf acquired within each time point (Figure 1a) is used for reconstructing one image so that L images are obtained and then used for template matching pixel by pixel. Since each interleaf is highly undersampled, the reconstructed images are seriously noisy due to the undersampling artifacts, leading to dramatic fluctuation in the signal evolution curve. To overcome the influence caused by that fluctuation on template matching, the number of time points (L) must be big enough to ensure the SNR in time domain.

In this work, a sliding window was utilized to select multiple interleaves, which are acquired during different time points, to generate one image for each step (Figure 1b). Therefore, the amount of reconstructed images became L-W+1, where W is the number of selected interleaves.

The result from combining spiral interleaves acquired with different repetition times and flip angles forms a multi-weighted image. If the window width is identical to the number of spiral rotating types, the reconstructed image would be a full-sample one. In this way, the random-like aliasing artifacts caused by undersampling error are reduced significantly, so are the fluctuations in signal evolution curve. Since the reconstructed images are multi-weighted, the dictionary has to be modified. Therefore, we performed a same sliding-window summation on original dictionary (DO) to generate a new dictionary (DN), namely $$$ \bf {DN}_\it{i}=\sum_{\it{j=i}}^{\it{i+W}}\bf {DO}_\it{j}$$$, to adapt to the signal evolution under new reconstruction method. The rest part is similar as the original MRF method.

To test the above strategy, we performed in-vivo brain experiment and phantom tests, using a MRF sequence based on an inversion-prepared FISP with TR varying from 10 to 12ms, flip angle varying from 5 to 80 degrees, and a variable density spiral (VDS) trajectory which was zero-moment compensated³. The spiral trajectory rotated 12 degrees for each TR so that the number of trajectory types was 30. Similarly, W was set as 30 to cover the full k-space. And the calculation of dictionary was based on extended phase graph algorithm⁴. The range of T1 in dictionary was from 20 to 6000ms, T2 from 20 to 3000ms.

A phantom of Polyvinylpyrrolidone (PVP) solution with concentrations from 10% to 60% was utilized for phantom validations. T1 and T2 values were also calculated by DESPOT1 and multi-TE TSE experiments respectively for comparisons. For in-vivo brain experiments, in addition to the conventionally acquired data, the reference parametric mappings were obtained by a full-sampled MRF sequence (L=1000 and 30 repetitions) with total scan time of 331s. All measurements were made on a Siemens Prisma 3T scanner.

For L=300 and L=500 (correspond to acquisition time of 3.4s and 5.6s, respectively), the phantom experiment (Figure 2) indicates that the results from proposed method were closer to reference values than the original method. And the in-vivo experiments (Figure 3) demonstrate that the sliding-window reconstruction had a better image quality with smaller NSSE (normalized sum-of-squares error) than the original method, especially for T2 mappings. Since the reconstruction of conventional MRF acquisition is based on single interleaf acquired from each time point, fluctuation in signal evolution curve is serious due to the undersampling artifacts. By using the sliding-window reconstruction strategy with a window width covering the period of VDS, the k-space of each reconstructed image is fully covered so that the fluctuation in signal evolution could be significantly alleviated. Therefore our proposed method can reach an expected result with reduced time points that leads to a faster acquisition.