Magnetic resonance fingerprinting (MRF) allows for the simultaneous quantification of multiple tissue properties via the matching of acquired signals to a pre-computed dictionary, created by sampling a wide range of the parameter space¹. As the parameters of interest increase, so does the dictionary size, leading to long reconstruction times. One possibility for overcoming this limitation is to use a clustered dictionary with both spatial and temporal information². This work aims at increasing MRF efficiency by using a clustered spatiotemporal dictionary and incorporating it into a MRF pipeline that includes B1 mapping and a view-sharing (VS) anti-aliasing strategy³.

We tested our approach using 3D MRF data of a Lister-hooded adult rat brain adult acquired with a Bruker BioSpec 47/40 system (Bruker Inc., Ettlingen, Germany)³. The sequence was based on SSFP-MRF⁴ with Cartesian sampling, T=1000 shots, and 0.5 mm isotropic resolution. A dictionary $$$D\in\mathbb{C}^{L \times T}$$$ was simulated using extended phase graphs with the following ranges: T1 from 100ms to 3,000ms in 20ms steps; T2 from 20ms to 100ms in 5ms steps and from 100 to 500ms in 10ms steps; and B1 as a flip angle factor from 50% to 150% in 1% steps, resulting in a dictionary of size 840522x1000. The acquired data was matched to the dictionary to create a reference dataset.

Exploiting symmetry of the brain, the reference dataset was divided along the medial longitudinal fissure, separating the left and right hemisphere. The estimated parametric T1, T2 and B1 maps of the left hemisphere (see Fig. 1) were used to create spatiotemporal dictionaries of different sizes by first clustering multi-parametric (T1,T2,B1) spatial kernels using k-means and subsequently simulating the temporal signal of every voxel in each cluster. The right hemisphere of the reference dataset was then matched to dictionaries with spatial kernel sizes of P=1x1x1 (clustered only), P=3x3x3 and P=5x5x5 (see Fig. 2).

We hypothesize that a dictionary that contains only feasible parameter combinations and spatial information should enable acceleration in both space and time. We test this by sampling less k-space points using a Gaussian mask in the phase encode directions with different acceleration factors (Figs. 3-4), and by reducing the sequence length (Fig. 4). Undersampled datasets were reconstructed with the original dictionary template matching (TM)¹ and with our VS approach, and compared to the reference dataset by their similarity index (SSIM)⁵. Furthermore, we study the amount of clusters required to accurately capture the entire spatio-parametric variability in our dataset by evaluating the mean square error (MSE) of the training and testing data for different spatial kernels (Fig. 5.).

Figure 1 shows how the estimated parameters approximate a Gaussian distribution, and are scattered in a restricted range within the parameter space. Hence, using dictionaries trained from this distribution yields parametric maps that agree with maps estimated using the full dictionary (see Fig. 2). Figure 3 compares the reconstructed maps with 20% sampling of k-space, where $$$D$$$ and $$$\hat{D}_1$$$ combined with VS are the most similar to the reference dataset. Figure 4 shows smaller variation of the clustered dictionaries with undersampling, though having less similarity to the reference dataset in fully sampled cases. Fig. 5 evidences how the training error decreases for more clusters in all cases, while the testing error only decreases continuously for $$$\hat{D}_1$$$. We use spatiotemporal dictionaries of different spatial kernel sizes with K=300 clusters (0.036% of the original dictionary size) and obtain comparable parametric maps (see Fig. 2). Furthermore, Figs. 3-4 show that clustered dictionaries, especially if they contain spatial information, are more robust to undersampling and shorter sequences. Conversely, the spatial smoothing achieved with larger spatial kernels along with the constant testing errors for increasing clusters in Fig. 5 indicate that the training data does not accurately represent the testing data for kernel sizes larger than P=3x3x3. In fact, the amount of training observations required and the corresponding size of the dictionary in terms of space, time, and clusters, leads to two important discussion points: 1) using clustering enables higher acceleration, at the expense of disregarding parameter combinations that are not present in the training set (e.g. pathology); and 2) adding spatial information increases the dimensionality of the dictionary, requiring approaches that can effectively deal with matching in high dimensional spaces. We propose a method to create clustered MRF dictionaries and show the added benefit of combining it with a view-sharing strategy to enable both accelerated acquisitions by undersampling, and accelerated reconstructions through dictionary compression. Further investigation of data-driven approaches could pave the way towards tissue and disease specific dictionaries in clinical settings.