CEREBRUM-7T

A 7-tesla scanner resolves cortical laminae that a 3T scanner cannot, but the field strength that buys the resolution also produces intensity inhomogeneity severe enough that the same tissue reads at different brightness in the frontal and occipital lobes. Tools built around 3T intensity statistics inherit that problem: FreeSurfer, even at v7, can fail to complete a fully automatic segmentation of a 7T volume, so most ultra-high-field sites fall back on in-house pipelines chaining bias correction, skull stripping, atlas registration and manual correction over several hours.

CEREBRUM-7T collapses that chain into one forward pass. Developed at the University of Glasgow with the University of Brescia and released as a preprint in July 2020 before publication in Human Brain Mapping in 2021, it is a fully volumetric 3D convolutional encoder-decoder that takes a raw out-of-the-scanner T1w MP2RAGE volume — 256 × 352 × 224 voxels, no atlas alignment, no inhomogeneity correction, no skull stripping — and returns a labelled volume in roughly ten seconds on a desktop CPU. Seeing the whole brain rather than tiles, it learns absolute and relative anatomical position as features, which patch-based segmenters discard.

The second idea is weak supervision. Manual annotation of 142 whole-brain volumes being infeasible, the authors trained against an inaccurate ground truth assembled from existing automatic tools, then tested whether the network exceeded its own teacher. It did: in a blinded comparison, expert neuroscientists preferred its masks to the labels it learned from.

#Key Features

• Whole-volume inference: The full brain is processed at once rather than as 2D slices or 3D tiles, preserving global spatial context that tiling loses at the seams.
• No preprocessing pipeline: The only step before inference is neck cropping. Bias-field handling and skull stripping are learned behaviours, not upstream tools.
• Weakly supervised beyond its labels: Training targets came from a pipeline combining AFNI-3dSeg, the method of Fracasso et al. (2016) and FreeSurfer v6, yet the learned model scores higher against manual annotation than any of those sources.
• Deliberately small: About 1.2M parameters, using strided convolutions for downsampling and tensor sums with skip connections rather than concatenation, which is what fits full-resolution volumetric processing into GPU memory.
• Documented transfer path: Pretrained weights ship with the code alongside a fine-tuning procedure for a different 7T site.

#Technical Details

The architecture is a three-level fully volumetric U-Net with 3 × 3 × 3 kernels, 24, 48 and 96 filters and one, two and three convolutional blocks per level, and a softmax over seven labels — background plus grey matter, basal ganglia, white matter, ventricles, cerebellum and brainstem, following the MICCAI MRBrainS labelling scheme. Training used 110 of 142 MP2RAGE volumes (0.63 mm isotropic, Siemens 7T Terra, 32-channel head coil) acquired in Glasgow, with 6 for validation and 26 held out, expanded tenfold by offline geometric augmentation and perturbed online with Gaussian, salt-and-pepper and a synthetic inhomogeneous-field noise that mimics 7T intensity drift. Optimisation was Adam at a learning rate of 5 × 10⁻⁴ on categorical cross-entropy, converging in about 23 epochs over roughly 24 hours across four GTX 1080 Ti cards. Evaluated against 2.7M manually labelled voxels spanning eight regions in three subjects, its Dice coefficient exceeded FreeSurfer v6, FreeSurfer v7, Nighres and the training labels on every structure. Fine-tuning on 20 automatically labelled volumes from the AHEAD database, or on three manually labelled volumes from an independent set, adapted the model to new sites, reaching total Dice of 0.977–0.980 in the latter case.

#Applications

The immediate consumers are laminar and high-resolution fMRI studies, where an accurate grey-matter boundary determines whether cortical-depth sampling is meaningful and a segmentation step measured in hours per subject constrains study size. The model outputs per-voxel probability maps as well as thresholded masks, so surface reconstruction can use the soft output directly. Its speed and determinism also suit reprocessing of shared 7T structural data on OpenNeuro and EBRAINS, where a fixed automatic tool removes the operator-to-operator variability that makes independently analysed cohorts hard to compare.

#Impact

CEREBRUM-7T was the first fully automatic deep-learning segmentation released for out-of-the-scanner 7T brain MRI, and the release was unusually complete for its moment: code, trained weights, Docker and Singularity containers, the 142-volume Glasgow database and its training labels on EBRAINS, a manually annotated subset on OpenNeuro, and the PsychoPy Turing test itself. Its principal limitation is the one its authors state — the model is trained on a single site's scanner and sequence, and moving it elsewhere is a fine-tuning operation on a reduced set of volumes rather than an out-of-the-box application, a constraint that motivated the group's later multi-site work, LOD-Brain. The stack has also aged: the checkpoint targets TensorFlow 1.14 and CUDA 10.0, so reproducing inference means rebuilding that environment or using the published containers.