Scientists interested in automatic segmentation of the thalamus and its nuclei, esp. for the study of atrophy or treatment targeting.The white matter nulled MP-RAGE (WMnMPRAGE) contrast at 7T has enabled detailed manual delineation of thalamic nuclei¹ by a trained neuroradiologist guided by the Morel atlas² as well as the creation of a fully automatic segmentation pipeline called THOMAS (Thalamus Optimized Multi-Atlas Segmentation), the first of its kind to segment the 12 thalamic nuclei at 1mm3 resolution with validation against expert-defined labels.³ However, the availability of 7T scanners is limited. In this work we investigate the application of THOMAS to WMnMPRAGE images acquired at 3T. The performance of the algorithm is evaluated over 5 subjects scanned at 3T and 7T against ground truth manual delineations from 7T images.Following informed consent, 6 essential tremor patients were scanned at 7T (Discovery MR950, GE Healthcare) using a 32-channel head coil (Nova Medical) with WMnMPRAGE scan parameters: TS 6s, TI 680ms, TR 10ms, BW 12kHz, flip angle 4°, FOV 18cm, 180x180x200 matrix, slice thickness 1mm, ARC parallel imaging 1.5x1.5 (2D radial fanbeam), scan time 5.5 minutes.⁴ Using these images, the whole thalami and 12 thalamic nuclei were manually delineated with a high degree of reproducibility to form the ground truth.1 WMnMPRAGE images were also acquired at 3T (Discovery MR750, GE Healthcare) using an 8-channel head coil: TS 4.5s, TI 500ms, TR 10ms, BW 9kHz, flip angle 7°, FOV 18cm, 180x180x200 matrix, slice thickness 1mm, no parallel imaging, scan time ~10 min with prospective motion correction. THOMAS was applied without any modification to all WMn images to automatically segment the thalamus and its nuclei. The pipeline involves non-linearly registering a library of prior thalamic atlases to the target subject and treating these as potential candidate segmentations. These candidates are combined together and their differences resolved with a label fusion by Wang and Yushkevich.⁵ The 7T images were registered to the 3T images using FSL’s FLIRT with 7 degrees-of-freedom. Special care was taken to ignore signal artifacts around the sinuses which were present at both field strengths but more severe at 7T in the registration. The results were evaluated against manually delineated truth labels using Dice’s coefficient. In the 3T case, the ground truth labels were transferred from 7T using the registration. One subject was removed from the analysis due to poor image quality from coil sensitivity fall off.Segmentation time for a single subject was 30-35 minutes on a 12-core 2.66 GHz Intel Xeon independent of field strength. Fig. 1 shows the performance of THOMAS for the whole thalamus and each of the 12 nuclei at 3T and 7T. Fig. 2 shows different in the dice between 3T and 7T. Notably, it appears that THOMAS works as well with 3T WMnMPRAGE as with 7T images. The difference in dice is generally within ±0.1 between the two on an individual subject basis (Fig. 2). Fig. 3 aggregates the Dice across all 12 nuclei and computes a kernel density estimate of the probability of segmentation performance at each field strength. A Wilcoxon signed-rank test has a p-value of 0.531, which indicates that the null hypothesis that the differences between the two are symmetric about 0 cannot be rejected. Example segmentations are depicted in Fig. 4.The THOMAS algorithm can successfully segment thalamic nuclei in both 3T and 7T WMnMPRAGE images with essentially equivalent quality, when evaluated against manual delineations. In one subject, THOMAS performed exceptionally well on the 3T images compared to 7T, we speculate that this may be due to improved B1+ uniformity at 3T. The initial step in the pipeline does bias-correct the images to remove any gross image non-uniformities but this may not be sufficient for 7T images and does not resolve non-uniformities in image contrast caused by B1+ non-uniformity. For these reasons, 3T WMnMPRAGE may eventually prove to be preferred over 7T WMnMPRAGE despite the improved signal-to-noise ratio at 7T. Reliable automatic segmentation of the whole thalamus and its nuclei is critically required for any large-scale study of thalamic changes with disease. It may also play an important role for treatment planning, e.g. in thalamic ablation for essential tremor or deep-brain stimulation for a variety of disorders. Demonstration that our method successfully operates at 3T means that it can find applicability in many more studies. Ongoing work will evaluate segmentation performance using a variety of 3T WMnMPRAGE volumes collected over a several year time period across different sites and scanner manufacturers.