This work presents a rapid field map prediction method based on the individual subject's quick localizer scan and a large brain field map database to accelerate the field map acquisition stage for dynamic shimming applications. The method is evaluated by comparison to a linear interpolated low-resolution field map (LRFM) and taking a high-resolution field map (HRFM) as a reference.Dynamic shimming has been demonstrated to provide a more homogeneous field than static shimming and reduces distortion and signal dropout in EPI.^1,2 However, additional time is required to acquire a high-resolution field map (~4 minutes) for calculating shim settings. Previously it has been shown that, instead of acquiring a high-resolution field map, it is possible to predict the subject-specific field inhomogeneity by incorporating brain structural information with a model of field distortion in the brain.³ This method was demonstrated to be capable of significantly shortening acquisition time compared with conventional field-map-based dynamic shimming, while outperforming static shimming. Here we introduce an optimized model-based field map prediction (MFMP) method by using a large brain field map database and k-mean clustering method to make the model more adaptive. To evaluate whether the model-based method is predicting field changes that cannot be estimated by a rapidly acquired, low-resolution field map, we compare both methods with a high-resolution field map.

The experiments were performed on a Siemens whole-body 7T MRI scanner (Erlangen, Germany). High-resolution (Voxel size=2mm*2mm*1.2mm, Slices=120, TR=885ms, TE1 =4.08ms, ΔTE=1.02ms, FOV=200mm*200mm) and low-resolution (Voxel size=5mm*5mm*5mm, Slices=30, TR=250ms) field maps were acquired using gradient echo (GRE) pulse sequences and used for slice-wise shim calculation. A field map database was generated based on 70 whole-brain field maps from different subjects acquired on the 7T scanner. Brain structural information used for shim model generation was acquired using a scanner default 3D localizing sequence (19 seconds).

In order to generate the template field map, the FLIRT (FMRIB's Linear Image Registration Tool)⁴ was used to register all field maps to a FSL structural template; the subject-specific field map was generated by taking the weighted average of the 5 most similar field maps, using a k-mean clustering method, in terms of Euclidian distance between the registration information (skew, scale, rotation) of the scanned subject and all the whole-brain field maps in database. Weightings (significantly to the skew parameters) were applied to bias the matching towards the lower brain where we know that predicting the field is most difficult and also most informative.

The mid-sagittal view of a subject's brain and the corresponding measured high-resolution field map are shown in Figure 2. Figure 3 compares field maps acquired using MFMP, interpolated LRFM and HRFM methods at the level of slice 32. Though both fields produced using MFMP method and interpolated LRFM can describe the general field inhomogeneity, the interpolated LRFM is less able to reproduce the field inhomogeneity in the frontal lobe.

Figure 4 shows the standard deviation of the residual fields after subtracting HRFM from field maps obtained using interpolated LRFM and MFMP methods. Lower standard deviation difference represents the predicted field is more close to the HRFM, which is a more accurate measure of field inhomogeneity and Figure 5 shows the difference between the MFMP and LRFP methods.

These figures show that the model-based method is less able to accurately map the field higher up in the brain, but there are some slices lower down in the brain that are better estimated by the MFMP method. It is unsurprising that there are regions where the model-based method cannot specifically predict the subject’s unique field profile, since there are many subject dependent variables, including the size of the body, that affect the shim and cannot be estimated on the basis of the localizer alone. However, the ability of the model-based method to better predict the steep field variations found in the inferior frontal lobes than the lower resolution field map, shows that this method has some utility.

Our model-based method is able to better identify the steep change in the field close to the sinuses in the lower part of the brain, however a low-resolution field map performs better for the rest of the brain. This shows that the model-based field map prediction method can provide additional information on the field inhomogeneities in individual subjects beyond that contained in a low-resolution field map. It would however benefit from the incorporation of low-resolution field map information in the upper part of the brain. A combination of the two techniques may therefore provide the optimum combination of prediction accuracy and acquisition speed.