MR-guided focused ultrasound (MRgFUS) brain systems deliver targeted thermal energy into the brain using a hemispheric array of transducers that surround the head with an intervening water bath (Fig 1a). During treatment the localized heating (hot spot) is measured from a change in image phase between baseline (pre-treatment) and dynamic (during treatment) images. Accelerating temperature mapping by undersampling k-space is desirable to increase spatiotemporal resolution and coverage, but is difficult to do with parallel imaging since coils must be placed outside the transducer, far away from the head. Multiple groups have instead developed accelerated temperature mapping methods that exploit temporal correlations between baseline and dynamic images [1,2]. However, circulation of the water bath to cool the skull causes dynamic signal changes that are not captured by baseline images (Fig 1b), which breaks those correlations and results in artifacts throughout the temperature maps. We propose a spatially-segmented approach for reconstructing temperature maps in brain MRgFUS, in which we separately estimate a water bath image without a baseline, and a temperature map in the brain with a baseline. The method can estimate artifact-free temperature maps from undersampled data during brain MRgFUS treatments using a single receive coil.Our iterative approach alternates between updating the parameters of a k-space hybrid signal model which is fit in the brain region of the image [1], and a baseline-free estimate of the water bath image. Fitting the k-space hybrid brain model results in a phase drift-corrected brain image without the temperature phase shift and a sparse temperature phase shift map. An algorithm to fit the model is described in [1]. The water bath is reconstructed using a POCS algorithm that alternately enforces data consistency, consistency with a water bath support mask (brain and water bath masks are obtained from a baseline image), and sparsity in the Coiflet domain using soft thresholding [3]. Figure 1c illustrates the overall undersampled dynamic image model.
To test the method, a gel-filled human skull phantom was sonicated by an Insightec ExAblate Neuro 4000 transcranial MRgFUS system (Insightec Ltd, Haifa, Israel) while imaging with a GE 3T MR750 scanner (GE Healthcare, Waukeshaw, WI). 27 single-slice gradient echo images were collected with the body coil and 28 x 28 x 0.3 cm³ field of view, 256 x 128 acquisition matrix, 30° flip angle, 13 ms TE, and 28 ms TR. Images and maps were reconstructed to a 128 x 128 matrix and retrospectively randomly undersampled by 2x, with full sampling over 22 central k-space lines. Temperature maps were reconstructed by fitting the k-space hybrid model to the entire image, or to the brain only with keyhole or POCS methods used to reconstruct the water bath image.Figure 1d shows the temperature reconstruction results. When the k-space hybrid model is fit to the entire image without distinguishing between the brain and water bath, phase artifacts obscure the hot spot in the reconstructed temperature map and (in this case) lead to an overestimation of the temperature rise in the sonicated region across image dynamics (RMSE across dynamics: 0.0121°C). Restricting the temperature reconstruction to within the brain, in combination with keyhole reconstruction of the water bath image (using the baseline image’s k-space to fill in missing k-space lines), produces temperature maps with lower errors in the hot spot but still large errors outside (RMSE across dynamics: 0.0039°C). The proposed k-space brain/POCS bath approach yields a more accurate estimate of the water bath image (not shown), resulting in much lower in-brain temperature artifacts (RMSE across dynamics: 0.0029°C).Unpredictable water bath motion confounds model-based approaches to accelerated MR temperature mapping, resulting in large temperature artifacts due to aliased water bath signal. We demonstrated that a spatially-segmented reconstruction that applies a model-based reconstruction in the brain and a POCS reconstruction in the water bath can reconstruct temperature maps without undersampling artifacts at a moderate acceleration factor using a single receive coil. Future work will focus on integrating the approach with other accelerated temperature mapping methods [2] and extending it to non-Cartesian trajectories [4]. The method is compatible with multiple receive coils.