GOUHFI 2.0: A Next-Generation Toolbox for Brain Segmentation and Cortex Parcellation at Ultra-High Field MRI

TL;DR

GOUHFI 2.0 tackles automatic brain segmentation and cortex parcellation in ultra-high field MRI by introducing two independently trained 3D U-Nets: a brain segmentation model covering 35 labels across various contrasts and field strengths, and a cortex parcellation model delivering 62 DK-Tourville labels, complemented by an integrated volumetry pipeline. The authors expand the training corpus with aged and demented subjects and employ domain randomization to achieve contrast- and resolution-agnostic performance on UHF data, addressing limitations of 3T-optimized tools. Results across multiple datasets show improved robustness in older and clinical cohorts, reliable cortical parcellations, and volumetry that aligns with standard workflows, while highlighting remaining challenges in brain extraction and TIV estimation. GOUHFI 2.0 thus provides a comprehensive, first-of-its-kind DL toolbox for cortex parcellation at UHF and supports wide-scale neuroimaging analyses across field strengths. Future work will focus on overcoming brain-extraction limitations and further enriching training with non-healthy anatomies to enhance TIV precision and generalizability.

Abstract

Ultra-High Field MRI (UHF-MRI) is increasingly used in large-scale neuroimaging studies, yet automatic brain segmentation and cortical parcellation remain challenging due to signal inhomogeneities, heterogeneous contrasts and resolutions, and the limited availability of tools optimized for UHF data. Standard software packages such as FastSurferVINN and SynthSeg+ often yield suboptimal results when applied directly to UHF images, thereby restricting region-based quantitative analyses. To address this need, we introduce GOUHFI 2.0, an updated implementation of GOUHFI that incorporates increased training data variability and additional functionalities, including cortical parcellation and volumetry. GOUHFI 2.0 preserves the contrast- and resolution-agnostic design of the original toolbox while introducing two independently trained 3D U-Net segmentation tasks. The first performs whole-brain segmentation into 35 labels across contrasts, resolutions, field strengths and populations, using a domain-randomization strategy and a training dataset of 238 subjects. Using the same training data, the second network performs cortical parcellation into 62 labels following the Desikan-Killiany-Tourville (DKT) protocol. Across multiple datasets, GOUHFI 2.0 demonstrated improved segmentation accuracy relative to the original toolbox, particularly in heterogeneous cohorts, and produced reliable cortical parcellations. In addition, the integrated volumetry pipeline yielded results consistent with standard volumetric workflows. Overall, GOUHFI 2.0 provides a comprehensive solution for brain segmentation, parcellation and volumetry across field strengths, and constitutes the first deep-learning toolbox enabling robust cortical parcellation at UHF-MRI.

Figures (9)

• Figure 1: Pipeline to create the training data for both 3D U-Nets used in GOUHFI 2.0. For all training cases, the T1w image is segmented using FastSurferVINN (FSV) at native image resolution. The label map is then used for two things. On one side, the CSF label is modified to include all unassigned brain voxels present inside the brain mask. This new label map is then fed into the generative model to create the domain randomized synthetic image and corresponding label map. This enables to create a dataset of synthetic images (A) for training a contrast agnostic 3D U-Net for brain segmentation. Separately, the cortex segmentation is extracted from the label map by masking all remaining labels. This cortex segmentation, with its corresponding parcellation, are fed into the same domain randomization model as (A) without incorporating signal inhomogeneity since these are label maps and not images. This results in a second training dataset (B) used to train a 3D U-Net to perform cortex parcellation from an input cortex segmentation.
• Figure 2: Training datasets used for both 3D U-Nets trained as part of GOUHFI 2.0. Dataset A corresponds to the training dataset used for the contrast-agnostic brain segmentation task as done for the original GOUHFI. One important detail for dataset A is that the training data is composed of synthetic images whereas the validation data is composed of real MR images in order to reflect the true usage domain with real MR images. Alternatively, dataset B consists only of discrete label maps where a single cortex segmentation label is used as training "image" and its corresponding cortex parcellation with 62 labels is the training label map.
• Figure 3: GOUHFI 2.0 processing pipeline. By using an image of any contrast, resolution and even field strength, one can obtain (1) a whole brain segmentation into 35 labels following the FreeSurfer lookup table and (2) cortex parcellation into 62 labels (31 per hemisphere) following the Desikan-Killiany-Tourville (DKT) convention. The volume of every structure in the output brain segmentation and cortex parcellation are calculated and exported for subsequent quantitative analyses. While not developed as part of GOUHFI 2.0, the reorientation and brain extraction steps are implemented in the toolbox, making it a completely standalone solution for brain neuroimaging analyses like FreeSurfer/FastSurferVINN or SynthSeg$^{+}$.
• Figure 4: Figure presenting the differences in segmentations between FastSurferVINN, SynthSeg$^{+}$, GOUHFI, GOUHFI 2.0-n1 and GOUHFI 2.0-n2, computed from an N4-corrected MP2RAGE image highly affected by signal inhomogeneities. The example is from a PDP subject in the STRAT-PARK dataset, exhibiting substantial ventricular enlargement. Red, yellow and green arrows point to the quality of the hippocampus-inferior lateral ventricle delineations across techniques. Blue arrows demonstrate the repeated improvement of the segmentation of enlarged ventricles throughout GOUHFI's versions. Orange arrows demonstrate the performance of the different algorithms regarding the segmentation of the cortex in presence of highly inhomogeneous signal.
• Figure 5: Figure demonstrating the difference in cerebellum segmentations between FastSurferVINN, SynthSeg$^{+}$, GOUHFI, GOUHFI 2.0-n1 and GOUHFI 2.0-n2 computed from the N4-corrected MPRAGE image for two healthy subjects in the SCAIFIELD dataset. For subject A, the green horizontal lines correspond to the most inferior cerebellar WM branches detected by GOUHFI 2.0-n2 for both hemispheres. The lines for GOUHFI 2.0-n2 are reproduced for every method for comparison. The orange arrows show errors in cerebellar WM segmentation. For subject B, red arrows show the mislabeling of the cerebellar cortex by GOUHFI 2.0-n1 compared to the other techniques. Yellow arrows compare the segmentation of one cerebellar WM branch between techniques.