Ultra high-field MRI enables sub-millimetre resolution imaging of the human brain, allowing for the resolution of functional circuits at the meso-scale of cortical layers. An essential step in many functional and structural neuroimaging studies is segmentation, the operation of partitioning the MR brain images to delineate anatomical structures. Despite recent efforts in brain imaging analysis, the literature lacks of accurate and fast methods for segmenting 7 Tesla (7T) brain MRI. We here present CEREBRUM-7T, an optimised end-to-end Convolutional Neural Network (CNN) architecture, that allows for the segmentation of a whole 7T T1w MRI brain volume at once, thus overcoming the drawbacks of partitioning the volume into 2D or 3D tiles. Training is performed in a weakly supervised fashion, exploiting labelling with errors obtained with automatic state-of-the-art methods. The trained model is able to produce accurate multi-structure segmentation masks on six different classes in only a few seconds. In the experimental part, a combination of objective numerical evaluations and subjective analysis carried out by experienced neuroimaging users, confirms that the proposed solution outperforms the training labels it was trained on in segmentation accuracy, and is suitable for neuroimaging studies, such as layer fMRI studies. Taking advantage of a fine-tuning operation on a reduced set of volumes, we also show how it is possible to efficiently and effectively apply CEREBRUM-7T to data from different sites. Furthermore, to allow replicability and encourage extensions, we release the code, 7T data (142 volumes), and other materials, including the training labels and the Turing test.


Below, reconstructed meshes of GM, WM, basal ganglia, ventricles, brain stem, and cerebellum of a testing volume, obtained with CEREBRUM7T on sub-013_ses-001 (mesh created with BrainVoyager). A light smoothing operation is performed (50 iterations) - no manual corrections.

In the following pages, there are results of our method with data from three different sites and settings:

Pros and Cons

Timing Very fast inference: ~ 5/10 sec. Training needs ~ 24 hours
Hardware Inference easily done on CPU Needs 4 GPUs for training
Training labels Does not need accurate labels Does not overcome systematic errors
Visual outcome Very clean and smooth segmentation Does not recover from "burn" artefacts
New dataset Works very well but only if... → Needs fine-tuning


Visit the relative page to learn how to use CEREBRUM-7T from source code, docker, or singularity.


Visit the relative page for all the information needed about the data.

Turing test

As we discussed in the paper, we designed and implemented a PsychoPy (Peirce et al., 2019) test for neuroscientists to assess the segmentation quality of CEREBRUM-7T comparing our model with manual segmentation and with the GT. In the figure below, we show the interface we presented the participants with.

The source code to replicate the test is made available here.


Svanera, M., Benini, S., Bontempi, D., & Muckli, L. (2020). CEREBRUM-7T: Fast and Fully-volumetric Brain Segmentation of 7 Tesla MR Volumes. bioRxiv.

@article {Svanera2020.07.07.191536,
	author = {Svanera, Michele and Benini, Sergio and Bontempi, Dennis and Muckli, Lars},
	title = {CEREBRUM-7T: Fast and Fully-volumetric Brain Segmentation of 7 Tesla MR Volumes},
	elocation-id = {2020.07.07.191536},
	year = {2020},
	doi = {10.1101/2020.07.07.191536},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2020/12/03/2020.07.07.191536},
	eprint = {https://www.biorxiv.org/content/early/2020/12/03/2020.07.07.191536.full.pdf},
	journal = {bioRxiv}


This project has received funding from the European Union Horizon 2020 Framework Programme for Research and Innovation under the Specific Grant Agreement No. 785907 and 945539 (Human Brain Project SGA2 and SGA3).