An open-source tool for fast segmentation of any brain MR scan with the NextBrain histological atlas
Poster No:
1639
Submission Type:
Abstract Submission
Authors:
Oula Puonti1, Jackson Nolan2, Robert Dicamillo2, Yael Balbastre3, Adria Casamitjana4, Matteo Mancini5, Eleanor Robinson3, Loic Peter3, Roberto Annunziata3, Juri Althonayan3, Shauna Crampsie3, Emily Blackburn3, Benjamin Billot6, Alessia Atzeni3, Peter Schmidt3, James Hughes3, Jean Augustinack2, Brian Edlow2, Lilla Zollei2, David Thomas3, Dorit Kliemann7, Martina Bocchetta3, Catherine Strand3, Janice Holston3, Zane Jaunmuktane3, Juan Iglesias2
Institutions:
1Danish Research Centre for Magnetic Resonance,, Copenhagen, Denmark, 2Martinos Center, Charlestown, MA, 3University College London, London, United Kingdom, 4Research Institute of Computer Vision and Robotics,, Girona, Spain, 5Cardiff University Brain Research Imaging Centre, Cardiff, United Kingdom, 6Massachusetts Institute of Technology, Cambridge, MA, 7University of Iowa, Iowa City, IA
First Author:
Co-Author(s):
Introduction:
Structural analysis of the brain at the subregion level (e.g., hippocampal subfields) has the potential to provide us with a more fine-grained understanding of aging in health (de Flores, 2015) and in diseases like Alzheimer's (Pereira, 2014). To facilitate such analyses, we have recently published NextBrain (Casamitjana, 2024), a histological atlas with 200 μm resolution and 333 regions of interest (ROIs). NextBrain includes a Bayesian segmentation tool with currently high computational demands that are impractical for most neuroimaging studies. Here, we present a new, much faster version of the tool (20 minutes vs. 2-3 days in multi-core workstations; or <5 minutes on a GPU). Our new tool segments in vivo and ex vivo scans of any resolution and MR contrast at 300 μm resolution, without parameter tuning, and is publicly available (github-pages.ucl.ac.uk/NextBrain).
Methods:
Bayesian segmentation alternates between estimating two sets of parameters: deformation of an atlas, and parameters of a Gaussian mixture (Ashburner, 2005). This iterative procedure is very slow at high resolution. Here, we propose a single-shot optimization scheme that capitalizes on recent developments in machine learning and accelerated nonlinear registration:
1) The input scan is segmented with SynthSeg (Billot, 2023), from which initial Gaussian parameters of 7 predefined tissue types are derived.
2) Using these statistics, we generate a synthetic scan that is spatially aligned with NextBrain but matches the contrast and resolution of the input.
3) The synthetic scan is aligned to the input with FireANTs (Jena, 2024) with a hybrid loss function combining the (coarse) SynthSeg labels and local cross correlation of image intensities.
4) The final Gaussian parameters are estimated with expectation maximization (as in Ashburner, 2005), which also yields the final segmentation.
1) The input scan is segmented with SynthSeg (Billot, 2023), from which initial Gaussian parameters of 7 predefined tissue types are derived.
2) Using these statistics, we generate a synthetic scan that is spatially aligned with NextBrain but matches the contrast and resolution of the input.
3) The synthetic scan is aligned to the input with FireANTs (Jena, 2024) with a hybrid loss function combining the (coarse) SynthSeg labels and local cross correlation of image intensities.
4) The final Gaussian parameters are estimated with expectation maximization (as in Ashburner, 2005), which also yields the final segmentation.
Results:
We demonstrate the flexibility and scalability of our method on two publicly available datasets: a 100 μm ex vivo scan (Edlow, 2019, resampled to 300 μm) and the OpenBHB dataset (3,227 T1-weighted scans at 1 mm resolution, Dufumier, 2022). We emphasize that, despite their vastly different contrast and resolution, the two datasets are segmented with the same command, without any parameter tuning.
Figure 1 shows an automated segmentation of the ex vivo scan, highlighting the detail achievable with our method on high-resolution data. When compared with the ground truth segmentation presented in (Casamitjana, 2024), our method achieved an average Dice score across ROIs equal to 0.621 ± 0.169, almost identical to that achieved by the original method (0.626 ± 0.180). On this high-resolution dataset, the run time was ~90 minutes using 8 cores; the original implementation took almost a week on the same machine.
Figure 2 shows the results on the in vivo dataset. Figure 2a-b shows a segmentation example, while Figure 2c shows a comparison between a coarse 36-class segmentation computed with a neural network trained on 1mm T1w scans (the baseline in Billot, 2023), and our automated segmentation, which labels clustered together to mimic the 36-class protocol. The correlation between the ROI volumes computed with the two approaches is very high (Figure 2c). The Dice scores comparing the segmentations directly were also very high for all ROIs. Finally, Figure 2d-f shows the ROI-wise correlation between age and volume, i.e., an "aging map" for our highly detailed parcellation. The results are consistent with those obtained with the original tools and presented in (Casamitjana, 2024). The run time on an 8-core machine was ~20 minutes.
Figure 1 shows an automated segmentation of the ex vivo scan, highlighting the detail achievable with our method on high-resolution data. When compared with the ground truth segmentation presented in (Casamitjana, 2024), our method achieved an average Dice score across ROIs equal to 0.621 ± 0.169, almost identical to that achieved by the original method (0.626 ± 0.180). On this high-resolution dataset, the run time was ~90 minutes using 8 cores; the original implementation took almost a week on the same machine.
Figure 2 shows the results on the in vivo dataset. Figure 2a-b shows a segmentation example, while Figure 2c shows a comparison between a coarse 36-class segmentation computed with a neural network trained on 1mm T1w scans (the baseline in Billot, 2023), and our automated segmentation, which labels clustered together to mimic the 36-class protocol. The correlation between the ROI volumes computed with the two approaches is very high (Figure 2c). The Dice scores comparing the segmentations directly were also very high for all ROIs. Finally, Figure 2d-f shows the ROI-wise correlation between age and volume, i.e., an "aging map" for our highly detailed parcellation. The results are consistent with those obtained with the original tools and presented in (Casamitjana, 2024). The run time on an 8-core machine was ~20 minutes.
·Figure 1
·Figure 2
Conclusions:
For the first time, our method allows segmentation of hundreds of ROIs from scans of any resolution and contrast with limited computational demands. The results show its accuracy and potential for large-scale volumetric analysis. Our tool is open-source and will enable researchers to increase the spatial granularity of their neuroimaging studies.
Modeling and Analysis Methods:
Bayesian Modeling
Image Registration and Computational Anatomy
Methods Development
Segmentation and Parcellation 1
Neuroinformatics and Data Sharing:
Brain Atlases 2
Keywords:
Aging
Atlasing
Machine Learning
Modeling
MRI
Open-Source Software
Segmentation
Spatial Normalization
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Provide references using APA citation style.
Billot, B. (2023). SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical image analysis, 86, 102789.
Casamitjana, A. (2024). A next-generation, histological atlas of the human brain and its application to automated brain MRI segmentation. bioRxiv.
Dufumier, B. (2022). OpenBhB: a large-scale multi-site brain MRI dataset for age prediction and debiasing. NeuroImage, 263, 119637.
Edlow, B. (2019). 7 Tesla MRI of the ex vivo human brain at 100 micron resolution. Scientific data, 6(1), 244.
de Flores, R. (2015). Structural imaging of hippocampal subfields in healthy aging and Alzheimer’s disease, Neuroscience, 309, 29-50.
Jena, R. (2024). FireANTs: Adaptive Riemannian Optimization for Multi-Scale Diffeomorphic Registration. arXiv preprint arXiv:2404.01249.
Pereira, J.B. (2014), Regional vulnerability of hippocampal subfields to aging measured by structural and diffusion MRI. Hippocampus, 24, 403-414.