Optimized fast dictionary matching for magnetic resonance fi ngerprinting based on echo-planar imaging for enhanced clinical workflow
Ingo Hermann1, Benedikt Rieger1, Jascha Zapp1, Sebastian Weingärtner1,2,3, and Lothar R. Schad1

1Computer Assisted Clinical Medicine, University Medical Center Mannheim, Heidelberg University, Mannheim, Germany, 2Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN, United States, 3Center for Magnetic Resonance Research, University of Minnesota, Minneapolis, MN, United States


In this work, an optimized fast group matching reconstruction for magnetic resonance fingerprinting based on echo-planar imaging was evaluted to enhance clinical usability. This scanner based 'on the fly' reconstruction reduced the reconstruction time by an acceleration factor of 10 shortening the reconstruction to 10 seconds. The fast group matching algorithm was tested in-vivo and compared with full dictionary matching and resulted in virtually no deviation in T1 and T2* maps facilitating the use of MRF in clinical routine.


Quantitative magnetic resonance imaging can be used to characterize tissue by its physical properties, providing quantiative biomarkers for various disease. A novel approach to quantify several tissue properties is magnetic resonance fi ngerprinting (MRF). Pseudo randomized pulse sequences generate various imaging contrasts which are compared to large simulated dictionaries specifying variations of tissue parameters. The reconstruction of parameter maps through dictionary matching of a measurement within multiple slices is time consuming, often beyond what is clinically acceptable. A group matching algorithm was proposed by Cauley et al.[1] to reduce post-processing times. Aim of this work is to implement a fast group matching reconstruction on the scanner platform for 'on the fly' reconstruction to enhance clinical usability.


A group matching algorithm is implemented for a MRF sequence based on echo-planar imaging (EPI)[2]. The group matching algorithm is based on Cauley et al.[1] where the full dictionary is divided into multiple small dictionaries to match a particular part of the dictionaries. Figure 1 illustrates the algorithm. A random dictionary element is chosen and correlated with the full dictionary containing over 100,000 elements (d), the best 10 matching entries (g) were merged in one group. The randomly chosen entry, which yielded more accurate results than the group mean value, is written in a look up table (LUT) defi ning the best matching values for the group. This is performed for all entries prior to the measurement starts. Around 10,000 groups (n) with their 10,000 group values are generated. For each slice a separate dictionary is created. After image acquisition the fi ngerprints of each pixel are correlated with the LUT (compare Fi g. 1) and the 10 % (k) best matching (closest related) groups are chosen for an entire correlation. This resulted in the best matching T1, T2* and B1+ values. Reconstruction on the scanner is implemented with C++ in ICE (Siemens Healthineers) and the group dictionaries are generated in MATLAB by Bloch simulations. The fast group matching algorithm is compared to pattern matching algorithm on the scanner and in MATLAB. For prove of concept the whole dictionary is matched with the group matching algorithm to determine deviations of the reconstruction. The reconstruction is tested in a phantom and in-vivo in 5 healthy and 5 multiple sclerosis subjects on a 3T Magnetom Skyra and Prisma.


A comparison of pattern matching and group matching reconstruction showed complete agreement with virtually no deviations in T1 and T2* for T1 < 300 ms and T2* < 20 ms with an acceleration factor of 10. Figure 2 shows the MRF maps of 5 dif ferent slices with group matching reconstruction. In Fig. 3 the absolute deviation of full reconstruction versus group matching is depicted in percentage for 1 and 100 matched groups (k). For 1 matched group the deviation is RSME = 2.3 % and for 100 matched groups RSME = 1.3 % respectively. With 1,000 matched groups there is no deviation (RSME = 0 %) when matching the dictionary with both algorithms. An acceleration factor of more than 10 was achieved for k = 1,000 leading to an entire reconstruction time for the generation of T1 and T2 maps for 60 slices on Prisma around 10 seconds and on Skyra around 30 seconds.


Dictionary matching as compared with conventional optimization schemes, bears the advantage of high robustness against local minima. The group matching algortihm provides the means, to exploit local similarities within the dictionary, to speed up the reconstruction while still sampling across the entire dictionary, therefore, maintaining the advantage of local minima robustness.

The group matching algorithm is a reliable and fast method for MRF acquisitions, and is compatible with MRF-EPI scans. Best results are obtained for a high number (10%) of small groups (~10 per group). Slower group sizes result in longer reconstruction times. For group sizes larger than 10, the acceleration rate is comparable to the number of entries per group. This provides means to achieve increased reconstruction fidelity by using larger dictionaries while maintaining clinically acceptable reconstruction times.


Full brain MRF mapping with 60 slices within 10 seconds was achieved, without loss in accuracy by the group matching algorithm on the scanner based reconstruction. The MRF sequence in combination with the group matching algorithm generated good quality T1, T2* maps and enabled clinical usability. No post processing algorithm is needed due to the implementation on the scanner, which facilitates the use of MRF in clinical routine.



[1] Cauley, S. F. et al. Fast group matching for MR fingerprinting reconstruction. Magnetic Resonance in Medicine 74, 523-528 (2015).

[2] Rieger, B., Zimmer, F., Zapp, J., Weingartner, S. & Schad, L. R. Magnetic resonance fingerprinting using echo-planar imaging: Joint quanti cation of T1 and relaxation times. Magnetic Resonance in Medicine 78, 1724-1733 (2016).


Figure 1: Illustration of group matching algorithm with the dictionary entries d, number of entries per group g, number of measured images m, number of groups n and number of matching groups k. A random entry is chosen from the dictionary and matched with the dictionary resulting in a group dictionary for n groups as a preprocessing step for the reconstruction. During the reconstruction, the look up table (LUT) is correlated with the measured signal and the best matching groups are taken. The measured signal is correlated with all entries of the best matching groups and the best matching value is taken.

Figure 2: T1, T2* and B1+ maps of five slices with the group matching algorithm of a healthy subject, reconstruction time of 10 seconds for whole-brain mapping with 60 slices.

Figure 3: Comparison of the group matching algorithm and the normal pattern matching. The baseline image is overlaid by the absolute deviation of group matching to normal matching T1, T2* and B1+ values. Values with zero absolute deviation are depicted as transparent. Top column shows the comparison for matching one group and bottom column the matching of 100 groups.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)