Fast segmentation with the NextBrain histological atlas

Fast segmentation with the NextBrain histological atlas | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Fast segmentation with the NextBrain histological atlas Oula Puonti , Jackson Nolan , Robert Dicamillo , Yael Balbastre , Adria Casamitjana , Matteo Mancini , Eleanor Robinson , Loic Peter , Roberto Annunziata , Juri Althonayan , Shauna Crampsie , Emily Blackburn , Benjamin Billot , Alessia Atzeni , Peter Schmidt , James Hughes , Jean Augustinack , View ORCID Profile Brian Edlow , Lilla Zöllei , David L Thomas , Dorit Kliemann , Martina Bocchetta , Catherine Strand , Janice Holton , Zane Jaunmuktane , Juan Eugenio Iglesias doi: https://doi.org/10.1101/2025.09.22.673638 Oula Puonti 1 Danish Research Centre for Magnetic Resonance, Centre for Functional and Diagnostic Imaging and Research, Copenhagen University Hospital – Amager and Hvidovre , Copenhagen, Denmark 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jackson Nolan 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site Robert Dicamillo 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site Yael Balbastre 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Adria Casamitjana 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom 4 Research Institute of Computer Vision and Robotics, University of Girona , Girona, Spain Find this author on Google Scholar Find this author on PubMed Search for this author on this site Matteo Mancini 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom 5 “Enrico Fermi” Research Center , Rome, Italy 6 Cardiff University Brain Research Imaging Centre, Cardiff University , Cardiff, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Eleanor Robinson 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Loic Peter 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Roberto Annunziata 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Juri Althonayan 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Shauna Crampsie 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Emily Blackburn 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Benjamin Billot 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom 7 Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology , Cambridge, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site Alessia Atzeni 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Peter Schmidt 8 Advanced Research Computing Centre, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site James Hughes 8 Advanced Research Computing Centre, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jean Augustinack 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site Brian Edlow 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States 9 Center for Neurotechnology and Neurorecovery, Department of Neurology, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Brian Edlow Lilla Zöllei 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site David L Thomas 10 Dementia Research Centre, Department of Neurodegenerative Disease, UCL Queen Square Institute of Neurology, University College London , London, United Kingdom 11 Neuroradiological Academic Unit, Department of Translational Neuroscience and Stroke, UCL Queen Square Institute of Neurology, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Dorit Kliemann 12 Department of Psychological and Brain Sciences, University of Iowa , Iowa City, IA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site Martina Bocchetta 10 Dementia Research Centre, Department of Neurodegenerative Disease, UCL Queen Square Institute of Neurology, University College London , London, United Kingdom 13 Centre for Cognitive and Clinical Neuroscience, Division of Psychology, Department of Life Sciences, College of Health, Medicine and Life Sciences, Brunel University of London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Catherine Strand 14 Queen Square Brain Bank for Neurological Disorders, Department of Clinical and Movement Neurosciences, UCL Queen Square Institute of Neurology, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Janice Holton 14 Queen Square Brain Bank for Neurological Disorders, Department of Clinical and Movement Neurosciences, UCL Queen Square Institute of Neurology, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Zane Jaunmuktane 14 Queen Square Brain Bank for Neurological Disorders, Department of Clinical and Movement Neurosciences, UCL Queen Square Institute of Neurology, University College London , London, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Juan Eugenio Iglesias 2 Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School , Boston, MA, United States 3 Department of Medical Physics and Biomedical Engineering, University College London , London, United Kingdom 7 Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology , Cambridge, MA, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: jiglesiasgonzalez{at}mgh.harvard.edu Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Structural brain analysis at the subregion level offers critical insights into healthy aging and neurodegenerative diseases. The NextBrain histological atlas was recently introduced to support such fine-grained investigations, but its existing Bayesian segmentation framework remains computationally prohibitive, particularly for large-scale studies. We present a new, open-source tool that dramatically accelerates segmentation using a hybrid approach combining: machine learning, contrast-adaptive segmentation; target-specific image synthesis; and fast diffeomorphic registration (all three with GPU support). Our method enables highly granular segmentation of brain MRI scans of any resolution and contrast ( in vivo or ex vivo ) at a fraction of the computational cost of the original method ( < 5 minutes on a GPU). We validate our tool on four different modalities ( in vivo MRI, ex vivo MRI, HiP-CT, and photography) across a total of approximately 4,000 brain scans. Our results demonstrate that the accelerated approach achieves comparable accuracy to the original method in terms of Dice scores, while reducing runtime by over an order of magnitude. This work enables high-resolution anatomical analysis at unprecedented scale and flexibility, providing a practical solution for large neuroimaging studies. Our tool is publicly available in FreeSurfer ( https://surfer.nmr.mgh.harvard.edu/fswiki/HistoAtlasSegmentation ). 1 Introduction Image segmentation is a fundamental component of human neuroimaging. Automatic delineation of anatomical structures on MRI scans is critical for quantifying volumes of regions of interest (ROIs), assessing their morphometry, and supporting downstream analyses such as functional localization or connectivity modeling. Structural segmentation facilitates the study of brain development ( Giedd et al., 1999 ; Gogtay et al., 2004 ), aging ( Bethlehem et al., 2022 ), and various neurological and psychiatric disorders (Jack Jr et al., 1999), and is routinely used to derive imaging biomarkers for conditions such as Alzheimer’s disease, schizophrenia, and multiple sclerosis ( Cuingnet et al., 2011 ; Dickerson et al., 2011 ; Frisoni et al., 2010 ). As the scale and diversity of neuroimaging datasets increase, robust and generalizable segmentation tools are needed to handle variation in acquisition protocols, scanner vendors, and patient populations. This is particularly important when sharing open-source tools (as opposed to developing in-house solutions), since few assumptions can be made on the MRI contrast properties of the input scans. For this reason, Bayesian segmentation methods based on generative models ( Ashburner & Friston, 2005 ; Pohl et al., 2006 ; Puonti et al., 2016 ; Van Leemput et al., 2002 ; Wells et al., 1996 ; Zhang et al., 2002 ) have long been central to widespread neuroimaging software pipelines such as SPM ( Friston et al., 1994 ), FreeSurfer ( Fischl, 2012 ), FSL ( Smith et al., 2004 ), or AFNI ( Cox, 1996 ). In terms of generalizability, a particularly interesting subclass of Bayesian segmentation methods uses unsupervised likelihood models, which make it adaptive to MRI contrast. Best represented by Unified Segmentation ( Ashburner & Friston, 2005 ) (distributed with SPM), these methods combine supervised anatomical priors (probabilistic atlases of anatomy) with unsupervised likelihoods that typically comprise a Gaussian mixture model (GMM) and a bias field model. Crucially, the parameters of the GMM and the bias field model are estimated directly from the input scan, which can thus have been acquired with any pulse sequence. Furthermore, because the models of anatomy and image formation are decoupled, high-resolution ex vivo data (e.g., histology, ex vivo MRI) can be leveraged to construct detailed atlases, which can subsequently be applied to the segmentation of markedly different, lower-quality in vivo images. We have successfully used ex vivo MRI and histology to build probabilistic atlases of the hippocampus ( Iglesias et al., 2015 ), amygdala ( Saygin et al., 2017 ), and thalamus ( Iglesias et al., 2018 ) at the subregion level, which can be flexibly applied to in vivo segmentation. Our group has recently extended ex vivo atlasing from a couple of brain structures to the whole human brain using histology ( Casamitjana et al., 2025 ). Our new atlas (“NextBrain”) comprises 333 ROIs per hemisphere, defined at 200 µm resolution. NextBrain includes a Bayesian segmentation framework capable of transferring this anatomical detail to in vivo MRI images. However, the original implementation is computationally intensive: segmenting a single scan at 300 µm resolution can take days on a standard CPU-based workstation. While GPU acceleration can reduce run time, it requires access to high-end hardware that is not available to many users, limiting the practicality of the tool for large-scale studies or clinical applications. One potential solution to this problem would naturally be deep learning, which provides fast inference times and impressive results in a variety of medical image segmentation tasks ( Isensee et al., 2021 ; Ronneberger et al., 2015 ), including neuroimaging ( Henschel et al., 2020 ; Pereira et al., 2016 ; Salehi et al., 2017 )). Compared with Bayesian segmentation, these approaches suffer from “domain shift”, i.e., they often struggle to generalize to unseen domains without careful finetuning or domain adaptation ( Guan & Liu, 2021 ; Wang & Deng, 2018 ). There have been attempts to address this domain shift problem; one representative example for brain MRI segmentation is SynthSeg ( Billot et al., 2023 ), which relies on domain randomization. Its design is directly inspired by Bayesian segmentation, and it relies on generating synthetic images with almost the same generative model. Importantly, these images are synthesized with random model parameters to simulate broad variations in image appearance, which leads to high robustness at test time. Despite these advances, current deep learning approaches remain limited in transferring high-resolution information from ex vivo domains to in vivo data, particularly when only sparse ground-truth annotations are available across a large number of ROIs – as in NextBrain. In this work, we present a fast and robust reimplementation of Bayesian segmentation that enables segmentation with NextBrain in practical times without specialized hardware. Our method leverages carefully chosen approximations, recent advances in deep learning for initialization, contrast-adaptive synthetic image generation, and fast diffeomorphic registration to deliver highly detailed segmentations of in vivo and ex vivo brain MR scans with minimal computational demands. Crucially, the pipeline runs without any manual parameter tuning and achieves performance comparable to the original implementation, while reducing run time by more than an order of magnitude. This contribution makes detailed brain parcellation accessible to a wider community and enables high-resolution neuroanatomical analysis at scale. 2 Methods 2.1 Bayesian segmentation and NextBrain Preliminaries Bayesian segmentation is typically formulated as an inference problem within a generative model of the data: where L denotes the label assignment (i.e., the segmentation), and I represents the observed image. In general, both the likelihood p ( I | L ) and the prior p ( L ) are governed by parameter sets, denoted θ I and θ L , respectively. Exact posterior inference requires marginalization over these parameters, which is intractable in practice. A common approximation is to compute point estimates of the parameters by maximizing their posterior distribution: where we have made the standard assumption that the priors over the two parameter sets are independent, i.e., p ( θ I , θ L ) = p ( θ I ) p ( θ L ). Given the estimated parameters, the final segmentation is computed by maximizing the posterior over labelings: Model Instantiation in NextBrain NextBrain ( Casamitjana et al., 2025 ) uses a generative model similar in structure to Unified Segmentation ( Ashburner & Friston, 2005 ) and SAMSEG ( Puonti et al., 2016 ), which we briefly summarize here. The prior is defined by a probabilistic atlas A , which specifies conditionally independent categorical distributions over K possible labels at each spatial location. This atlas is deformed using a nonlinear transformation ϕ , parameterized by θ L . The deformation follows a probabilistic model encoded by a prior distribution p ( θ L ), which penalizes excessive geometric distortion via a membrane energy term. Specifically, the probability distribution over the labeling L is defined as: where j indexes the J voxels in the image, and ( x j , y j , z j ) represents the spatial coordinates of voxel j . The prior over the deformation parameters θ L is given by: where λ is a regularization constant, and E mem denotes the membrane energy, defined as: Given a labeling L , the likelihood models the image intensities using a GMM, assuming conditional independence across voxels and incorporating a smoothly varying multiplicative bias field. To simplify the model, the image intensities I are log-transformed, so that the bias becomes additive and the need to rescale probability densities due to multiplicative bias is eliminated. The likelihood is given by: where g ( · ; µ, σ ) denotes the probability density function of a Gaussian distribution with mean µ and standard deviation σ , and are the coefficients of the bias field basis functions . The full set of likelihood parameters is given by θ I = {µ 1 , …, µ K , σ 1 , …, σ K , c 1 , …, c B } , which includes class-specific means and variances, along with the bias field coefficients. A non-informative (uniform) prior is assumed over θ I , such that no specific MRI contrast of bias field is favored: p ( θ I ) ∝ 1. Segmentation as Bayesian inference Substituting the prior and likelihood models ( Equations 3 , 4 , and 5 ) into Equation 1 , taking logarithm, and disregarding constant terms that do not affect the optimization, the following objective function is obtained: This objective function is maximized using a coordinate ascent scheme, in which the parameters θ L and θ I are alternately optimized. The prior parameters θ L (representing the atlas deformation) are updated using standard numerical optimization methods ( Nocedal & Wright, 2006 ). The likelihood parameters θ I are optimized using the generalized Expectation Maximization (GEM) algorithm ( Dempster et al., 1977 ). GEM alternates between constructing a lower bound on the objective ℒ via Jensen’s inequality, and maximizing this bound with respect to θ I . This iterative process guarantees an increasing sequence of values for the original objective ℒ. Maximizing the lower bound involves iteratively updating the means, variances, and bias field parameters. All of these can be updated in closed form given the others. After convergence of this coordinate ascent procedure, the estimated parameters and are used to compute the final segmentation, as described in Equation 2 . In practice, explicit computation of this equation is unnecessary, since voxel-wise posterior probabilities over labels (also known as “responsibilities”) are already computed as part of the GEM algorithm. These responsibilities are used in the E-step to construct the lower bound. For further details on this framework, the reader is referred to Ashburner and Friston, 2005 or Puonti et al., 2016 . Computational bottlenecks The original inference algorithm in NextBrain is computationally intensive and often impractical for routine use. Its complexity arises from a deeply nested optimization structure at three levels: (i) alternating between deforming the atlas and updating intensity model parameters; (ii) the latter uses a GEM algorithm that iterates between E and M steps; and (iii) the M-step itself involves iterative updates of class means, variances, and bias field coefficients, each depending on the others. NextBrain’s nested optimization structure results in long run times, especially when combined with its large number of anatomical labels and high spatial resolution. NextBrain includes 333 ROIs per hemisphere, which are coarsely grouped during inference but ultimately expanded to the full label set when computing the final segmentation. The atlas is defined at 0.2 mm isotropic resolution, i.e., 125 times more voxels than a standard 1 mm MRI. Even at the default working resolution of 0.3 mm, voxels are 37 times smaller. Although resolution could be further reduced to alleviate computational requirements, this would compromise the benefits of a high-resolution histological atlas. The combination of model complexity, label granularity, and high spatial resolution results in substantial computational and memory demands. Segmenting a single case at 0.3 mm resolution takes over 48 hours on a CPU and still requires about 60 minutes on a GPU with at least 24 GB of memory. Ex vivo segmentation at 0.2–0.3 mm resolution is often infeasible with the current implementation. This large computational footprint limits the practicality of NextBrain as a segmentation tool. In the next section, we present a reimplementation that preserves the advantages of using our high-resolution atlas while dramatically improving run time and hardware efficiency through targeted approximations and algorithmic redesign. 2.2 Approximate inference for fast segmentation Overview Here we present a fast segmentation method for NextBrain that efficiently approximates the inference problem defined in Equation 1 . Unlike traditional approaches, our algorithm visits each set of model parameters only once, using GPU-enabled machine learning and a custom GPU-accelerated diffeomorphic registration algorithm. Importantly, the method preserves robustness to variations in MRI modality and contrast by maintaining contrast-adaptive modeling of image intensities. Although the registration involves several hyperparameters, our publicly available implementation uses a fixed set of defaults (employed in all experiments) that do not require manual tuning. Preprocessing The input scan is first segmented using the segmentation module of our foundation model “BrainFM” ( Liu et al., 2025 ). This method closely follows SynthSeg ( Billot et al., 2023 ), and produces coarse segmentations at 1 mm resolution that are highly robust to variations in contrast, resolution, and imaging artifacts (including strong bias fields). Compared to the original SynthSeg model, BrainFM adds several labels, including five contralateral ROIs in the limbic system ( Greve et al., 2021 ) and 12 extracerebral structures (see Liu et al., 2025 for details). These coarse segmentations are used to exclude non-brain tissue, including cerebrospinal fluid (CSF), which NextBrain models with the same class as the background. The segmentation is also used to divide the brain into hemispheres. Non-hemispheric ROIs such as the brainstem and white matter lesions are also split by thresholding the left-right component of a deformation field mapping the scan to a symmetric MNI atlas; this deformation is computed within BrainFM using the Registration-by-Regression (RbR) algorithm ( Gopinath et al., 2024 ; Hu et al., 2025 ). In addition to preprocessing, the ROI segmentations also play a key role when estimating model parameters in later stages of the pipeline, as described below. Optimizing the bias field parameters {c b } In practice, Bayesian segmentation algorithms group ROIs into a smaller number of tissue classes during inference, clustering regions expected to share similar intensity profiles (e.g., gray matter structures such as the cerebral cortex, hippocampus, and amygdala). This strategy improves robustness, especially for small regions that may otherwise yield noisy parameter estimates. For instance, since the third ventricle is a thin layer of cerebrospinal fluid, its intensity distribution can be more reliably estimated when grouped with the much larger lateral ventricles. Grouping also benefits bias field estimation by signaling that some ROIs should have similar intensity profiles, despite potentially being distant from each other (e.g., left and right putamen); this information greatly helps disambiguate intensity variations due to bias field and anatomy. In our proposed method, we model the bias field with a 3D discrete cosine transform (DCT) basis of order up to 6, resulting in 343 basis functions. To estimate the bias field parameters {c b } , we first extract the whole brain using the available segmentation and group the BrainFM labels into eight tissue classes: cerebral gray matter, cerebellar gray matter, cerebral white matter, cerebellar white matter, brainstem, cerebrospinal fluid (CSF), thalamus, and pallidum. A soft segmentation is then generated by blurring the one-hot encoding of this 8-class map with a Gaussian kernel ( σ = 0.5 mm ). This soft segmentation is used to perform a single M-step of the GEM algorithm, with responsibilities fixed to the soft assignments. Due to the high reliability of the SynthSeg-based segmentation, the large number of voxels available, and the relatively low number of parameters, reclassifying voxels in an EM loop offers negligible benefit. While the M-step entails a nested optimization that alternates between estimating tissue class statistics (means and variances) and updating the bias field coefficients, convergence is rapid. Optimizing the prior parameters (atlas registration) θ L In Bayesian segmentation, updating the atlas deformation corresponds to optimizing the objective in Equation 6 with respect to the deformation parameters θ L , while holding the intensity model parameters θ I fixed. This step amounts to a deformable image registration problem, where the data term is given by the log-likelihood of the image under the atlas and the current estimate of θ I . However, this registration becomes computationally expensive in the classical coordinate ascent scheme, as every θ I update effectively changes the objective function with respect to θ L , making the optimization slow. To avoid this costly alternation, we leverage the available BrainFM segmentation to bypass the need to update θ I iteratively. Despite its lower resolution compared to the NextBrain atlas, the BrainFM segmentation is sufficiently accurate to estimate class-wise means and variances from the bias field corrected image. We use these parameters to synthesize a Gaussian image (a “cartoon”) from the probabilistic atlas, which is then resampled to match the resolution of the input scan ( Figure 1 ). This procedure ensures that the cartoon exhibits contrast and resolution closely resembling those of the target scan, thereby greatly facilitating registration in the subsequent steps described below. Download figure Open in new tab Figure 1: (a) Coronal and sagittal slices of an ex vivo scan at 0.2 mm resolution ( Edlow et al., 2019 ). (b) Matching cartoon after registration with FireANTs and composite similarity metric. (c) Slices of a 1 mm isotropic, skull-stripped, T1-weighted scan from the MIRIAD dataset ( Malone et al., 2013 ). (d) Registered cartoon. Note the differences in contrast and resolutions of the cartoons (c,d), which attempt to match the scans they seek to segment. Because the NextBrain atlas includes a large number of ROIs, we group them into 16 tissue classes to generate a sufficiently detailed cartoon: cerebral white matter, cerebral gray matter, cerebellar white matter, cerebellar cortex, caudate, putamen, pallidum, lateral thalamus, medial thalamus, red nucleus, compact brainstem white matter, diffuse brainstem white matter, hypothalamus, mammillary bodies, dentate nucleus of the cerebellum, and hippocampal white matter. The mean intensities for these groups are derived from the BrainFM segmentation when a direct label correspondence exists (e.g., cerebral white matter); otherwise, we use heuristic estimates based on linear combinations of gray and white matter means (see Supplementary Material for details). We perform registration between the synthetic cartoon and the input image one hemisphere at a time, as NextBrain is a single-hemisphere model. Registration is performed using FireANTs ( Jena et al., 2024 ), which we run in greedy, non-symmetric mode (i.e., with the loss computed in the fixed image grid). The similarity metric comprises two terms: (i) local normalized cross-correlation (LNCC, Avants et al. 2008 ), computed with a kernel radius of 3 voxels; and (ii) the mean Dice score between the BrainFM segmentation and the warped atlas labels, where the atlas ROIs are grouped to match the BrainFM protocol. The optimization is carried out using the Adam optimizer ( Kingma & Ba, 2014 ), and proceeds in a multiscale fashion, from coarse to fine resolution levels. While FireANTs does not explicitly include a regularization term in the objective function, smoothness is implicitly enforced through the multiscale optimization scheme and the use of diffeomorphic transformations parameterized in a smooth velocity field space. This setup discourages implausible deformations by naturally limiting high-frequency updates, ensuring anatomically consistent mappings without the need for hand-tuned regularization weights. We used Gaussian kernels with σ = 0.25 and σ = 1.0 to smooth the field and its gradient, respectively. Examples of brain scans of different contrast and resolution and their corresponding registered cartoons are shown in Figure 1 . Optimizing the Gaussian parameters and obtaining the final segmentation Given the estimated bias field and atlas deformation parameters, the class means and variances can be optimized using the EM algorithm; we note that it is EM and not GEM because the fixed bias field enables global maximization of the lower bound at each iteration. The EM procedure is carried out at a predefined resolution, to which both the input image and the probabilistic atlas are resampled. In our implementation, we set this resolution to 0.4 mm for in vivo scans, which balances anatomical detail and computational efficiency. For ex vivo scans, we set it to 0.2 mm (the resolution of the atlas) or the resolution of the scan, whichever is lower. EM inference is performed over the same 16 tissue classes used previously to generate the cartoon for registration. Consequently, the soft assignments (responsibilities) from the final E-step yield a probabilistic segmentation into 16 tissue classes rather than fine-grained ROIs. Posterior probabilities over the full set of ROIs are obtained by assigning each ROI a fraction of the posterior equal to the ratio between its prior probability and that of the associated tissue class ( Puonti et al., 2016 ). Other simplifications and practical details To further improve the efficiency and practicality of our method, we implemented some additional algorithmic and engineering simplifications: As mentioned above, we segment images at 0.4 mm isotropic resolution, instead of the 0.3 mm resolution used in the original NextBrain implementation. This reduces computational and memory demands by a factor of 2.4 while preserving anatomical detail. The entire pipeline, including FireANTs, is implemented in PyTorch, enabling seamless deployment on both GPU and CPU. PyTorch also facilitates efficient multi-threaded CPU execution. To accelerate atlas deformation, we store the probabilistic atlas as a set of sparse vectors defined within tight bounding boxes for each ROI, which allows rapid deformation of only the relevant subregions. In practice, GPU memory usage is dominated by the estimation of likelihood parameters, whereas actual computation is dominated by the atlas registration. When processing high-resolution ex vivo images at resolutions finer than 0.4 mm, we provide the option to offload likelihood optimization to the CPU (where memory is typically more abundant) while keeping the registration step on the GPU. To further reduce run time and memory usage, we simplified the original NextBrain atlas by removing the smallest ROIs. The final model includes 264 anatomical regions instead of 333; the list is provided in Table S2. On modern hardware, a full case (both hemispheres) processes in approximately 5 minutes on a GPU, or 20 minutes on a multi-core CPU workstation. 3 Experiments and Results 3.1 Data We use five different datasets to evaluate the proposed method: Ex vivo MRI This dataset consists of 21 specimens imaged with a Siemens 7 Tesla scanner at 0.12 mm isotropic resolution. The average age of the cohort is 65 years and the samples are from donors without neurological disorders. The samples comprise 12 single left hemispheres, 7 single right hemispheres, and two full brains. One of the full brains corresponds to the case described in Edlow et al., 2019 , for which we manually annotated 98 ROIs (including subcortical regions) in the right hemisphere at 0.2 mm isotropic resolution; these labels were released as part of the NextBrain dataset ( Casamitjana et al., 2025 ). No gold-standard segmentations are available for the rest of the cases. In vivo MRI (OpenBHB) OpenBHB ( Dufumier et al., 2022 ) is a public meta-dataset with ∼ 1 mm isotropic T1-weighted volumes of 3,227 healthy subjects scanned at over 60 sites. The age range of the subjects is 6-86 years (mean: 25.2 years). No ground-truth segmentations are available for these scans. In vivo MRI (MIRIAD) The Minimal Interval Resonance Imaging in Alzheimer’s Disease (MIRIAD) ( Malone et al., 2013 ) is a public dataset with ∼ 1 mm isotropic T1-weighted volumes of 69 subjects (46 mild–moderate Alzheimer’s subjects and 23 controls, age at entry 69.5 ± 7.1 years, 31 men). It comprises 708 scans acquired with the same scanner and pulse sequences at intervals of 2, 6, 14, 26, 38 and 52 weeks, 18 and 24 months from baseline, with up to 12 scans per individual. Out of the 708 scans, 185 are repeat acquisitions of the same timepoints, which can be used to assess test-retest reliability. No ground-truth segmentations are available for these scans. Ex vivo Hierarchical Phase-Contrast Tomography (HiP-CT) HiP-CT is an advanced X-ray imaging technique used to visualize intact human organs ex vivo at multiple scales, from whole-organ structure down to cellular detail ( Walsh et al., 2021 ). It leverages phase contrast generated by X-ray wavefront distortions using a synchrotron source to achieve resolutions as high as 1 µ m. Walsh et al., 2021 have made 6 brain HiP-CT scans available (4 males, 2 females, ages 63-80). The scans were downloaded from the Human Organ Atlas (HOA) repository at a resolution that was closest to 0.3 mm isotropic, which resulted in a range of resolutions spanning from 0.2–0.3 mm. We excluded case LADAF-2020-31, as it was among the earlier acquisitions and its image quality was notably inferior to that of the other cases. No ground-truth segmentations are available for these scans. Ex vivo Visible Human 2.0 This is a high-resolution anatomical dataset of the head and neck of a 66-year-old male donor, focused on axial cryosection photography acquired every 0.15 mm ( Ratiu et al., 2003 ). Tissue contrast was optimized to distinguish fine brain structures. We normalized the photographs by median intensity to correct for illumination inhomogeneity across slices. 3.2 Results 3.2.1 Segmentation of ultra-high resolution ex vivo MRI Figure 2 presents segmented slices from the the case described in Edlow et al., 2019 , including both the manually labeled ROIs and the corresponding automated segmentations. Despite the difference in granularity between the two approaches (98 manually defined ROIs versus 264 in the automated output), the overall agreement between them is remarkably strong. Table S1 in the Supplementary Material summarizes the Dice scores for the 98 manually labeled ROIs, obtained by clustering the automated segmentation to align with the gold standard definitions. ROI sizes are also included in the table. As expected, there is a clear association between region size and Dice score. Larger structures (e.g., cerebral and cerebellar white matter and cortex) achieve Dice values near 0.9. Smaller regions tend to show lower scores, although only a handful fall below 0.4, which remains adequate for localization purposes. Overall, the Dice performance (0.610 ± 0.207) closely matches that of the original method (0.626 ± 0.180), indicating comparable segmentation accuracy. Download figure Open in new tab Figure 2: Axial (top row), coronal (middle), and sagittal slices (bottom) of the publicly available ex vivo scan from Edlow et al., 2019 , along with manual delineations (only right hemisphere available) and the automated segmentation with NextBrain. We note that the input was resampled from 0.1 mm to 0.2 mm resolution, which is the voxel size of the manual segmentation, and also the native resolution of NextBrain. We also note that the manual delineation is less granular than the automated segmentation (98 ROIs vs 264). Qualitative results are shown in Figure 3 , which displays sample coronal segmentations of five other subjects from the ex vivo MRI data set. As no reference segmentations are available for 20 out of the 21 subjects, we instead report the volumes of the substructures of the hippocampus, thalamus and amygdala in tables in Tables 1 , 2 , and 3 , and compare them with those from other existing ex vivo atlases ( Iglesias et al., 2015 , 2018 ; Saygin et al., 2017 ). Because these atlases were constructed from manual delineations based on anatomical protocols that differ from those of NextBrain, we mapped each label in the atlases to the closest corresponding NextBrain ROI or set of ROIs. The volumes and ROI correspondences are listed in Tables 1 , 2 , and 3 . In general, the volumes agree quite well, with two exceptions: small or narrow structures like the granular layer of the dentate gyrus in the hippocampus; and the posterior regions of the thalamus, due to discrepacies in manual delineation protocols – in Iglesias et al., 2018 , the boundary between mediodorsal and pulvinar regions is more anterior than in NextBrain, and the whole thalamus is also generally larger (6,540 vs 5,970 mm 3 ). View this table: View inline View popup Download powerpoint Table 1: Average volumes of the hippocampal subfields from Iglesias et al., 2015 (N=15) and from the automated NextBrain segmentations on the ex vivo MRI data (N=20). View this table: View inline View popup Download powerpoint Table 2: Average volumes of nuclei of the amygdala from Saygin et al., 2017 (N=10) and from the automated NextBrain segmentations on the ex vivo MRI data (N=20). View this table: View inline View popup Download powerpoint Table 3: Average volumes of nuclei of the thalamus from Iglesias et al., 2018 (N=6) and from the automated NextBrain segmentations on the ex vivo MRI data (N=20). Download figure Open in new tab Figure 3: Five example subjects from the ex-vivo MRI data set. Coronal view of the MRI data (top) and the corresponding automated segmentation using NextBrain (bottom). 3.2.2 Segmentation of in vivo MRI Dice and volume correlation on OpenBHB As with most human brain MRI segmentation tools, one of the primary applications of NextBrain is the analysis of in vivo scans acquired at near 1 mm isotropic voxel size. However, evaluating subregion-level Dice scores on such data would require manual annotations of structures that often cannot be reliably delineated at this resolution. Instead, we assess performance at the whole-ROI level by merging the automatically segmented subregions to approximate the FreeSurfer protocol ( Fischl et al., 2002 ). As a reference, we used the “supervised baseline” U-Net from Billot et al. (2023) , trained on 1 mm isotropic T1-weighted scans, to segment the 3,227 subjects from the OpenBHB dataset. While using automated segmentations as ground truth constitutes a silver (rather than gold) standard, this approach enables evaluation over a much larger and more diverse sample. The heterogeneity of OpenBHB further strengthens the generalizability of the assessment. Figure 4 illustrates the segmentation results for a representative case from the OpenBHB dataset. Our method produces segmentations that are qualitatively very similar to those of the baseline, while also leveraging visible subregional boundaries and additional anatomical labels to more precisely align the fine-grained NextBrain labels with the underlying image data. Noticeable improvements include: clearer separation between the putamen and adjacent cortex; subdivision of the thalamus that reflects the darker contrast of medial and posterior nuclei; more accurate delineation of the lateral boundary of the thalamus; subdivision of the pallidum; and improved characterization of the hypothalamic and subthalamic regions, among others. Download figure Open in new tab Figure 4: Segmentation of sample subject (368981431172) from OpenBHB dataset. (a) Axial slice of input. (b) Reference segmentation with supervised U-Net for 1 mm T1 scans. (c) Segmentation using NextBrain, with ROI boundaries highlighted in red. (e-f) Coronal slice of the same case. Table 4 presents the quantitative results for both Dice scores and volume correlations. There is a strong correlation between ROI volumes computed by the two methods. Dice scores comparing the segmentations directly are also consistently high across all regions. These findings support the robustness of the segmentation approach at the whole-region level. View this table: View inline View popup Download powerpoint Table 4: The average Dice scores and volume correlations against the silver standard on the OpenBHB data set. The NextBrain ROIs are grouped to match the coarser segmentation from the supervised U-Net. Figure 5 illustrates the ROI-wise correlation between age and volume (an “aging map”) based on our high-resolution parcellation. The results closely align with those reported using the original tools in Casamitjana et al., 2025 , revealing region-specific patterns of age-related volume change. Notable effects include stronger associations in the frontal cortex, the medial thalamic subregions, the anterior caudate, and the subicular region of the hippocampus. Download figure Open in new tab Figure 5: Slices of map of Spearman correlation for ROI volume vs age using OpenBHB dataset, restricted to subjects aged 35 years and over ( n = 431 subjects, mean age: 57.9 years); we chose this age range as 35 is approximately the age when age-related atrophy begins. Test-retest reliability with MIRIAD When processing in vivo MRI scans at approximately 1 mm resolution, NextBrain necessarily relies more heavily on geometric priors encoded in the atlas to delineate many internal boundaries – some of which are more reliably inferred than others from the outer contours of brain regions. It is therefore important to evaluate the test-retest reliability of the method. For this purpose, we used the MIRIAD dataset, which has 185 timepoints with repeat acquisitions. We computed the intra-class correlation coefficient (ICC) for the volumes of all 264 ROIs, averaged across the left and right hemispheres. An complete list is provided in Table S2 of the supplementary material. The vast majority of ROIs exhibit excellent reliability, with ICC values exceeding 0.8. Notably, all ROIs with lower ICCs are very small in volume ( Figure 6 ): all with ICC < 0.9 have a volume below 170 mm 3 (except the myelencephalon), and all with ICC < 0.8 fall below 30 mm 3 . Download figure Open in new tab Figure 6: Scatter plot of ICC vs average ROI volume in test-retest experiment on MIRIAD dataset. 3.2.3 Segmentation of unconventional, high-resolution ex vivo modalities Figure 7 shows coronal slices of all five HiP-CT scans along with the segmentation produced by NextBrain. Notably, our method can tackle significant variations from normal appearing anatomy as exemplified by the three subjects with substantially enlarged or deformed ventricles. Figure 8 shows orthogonal slices from the Visible Human dataset alongside the corresponding segmentation. Given the higher native resolution of this dataset compared to NextBrain, the segmentation was performed at half resolution (0.3 mm isotropic) to reduce computational load. The segmentations were computed on grayscale images; although our method naturally extends to RGB inputs (similar to Bayesian segmentation, with mean vector and convariance matrices replacing scalar means and variances), processing full color volumes at this resolution would require prohibitive amounts of memory. The high-resolution and excellent contrast of both ex vivo data sets allow for accurate delineation of complex structures such as the thalamic nuclei. The Visible Human data has increased detail in regions that are often difficult to visualize in MRI, e.g., the dentate gyrus of the cerebellum, as seen in the sagittal slice. Download figure Open in new tab Figure 7: Coronal slices of the five HiP-CT scans (top) and their corresponding automated segmentations using NextBrain (bottom). Download figure Open in new tab Figure 8: Coronal (left), axial (middle), and sagittal slices (left) of the Visible Human 2.0 dataset, along with automated segmentation computed with NextBrain. 4 Discussion and conclusion In this work, we introduced a fast segmentation method based on the NextBrain histological atlas, capable of accurately parcellating a large number of brain ROIs across several modalities. A key strength of the proposed approach is its computational efficiency, which enables practical application to high-resolution datasets without incurring prohibitive computational costs. For example, segmenting the Visible Human volume at 0.3 mm isotropic resolution required only ∼ 90 minutes on a standard eight-core workstation, compared to nearly a week using the original implementation. For in vivo scans, our method completes segmentation in approximately 20 minutes on the same hardware, whereas the original code required 2–3 days. When a GPU with sufficient memory ( ∼ 24 GB) is available, segmentation time can be reduced to under five minutes. Quantitative validation across several benchmarks, including manual annotations, volume comparisons, and comparison against established segmentation methods, demonstrated that our approach achieves competitive accuracy. As expected, segmentation performance is positively correlated with ROI size, but even small regions such as thalamic subregions, hippocampal subfields, and subregions of the amygdala were reliably identified. Moreover, the method captures biologically meaningful variation: age-volume correlations derived from the segmentations revealed region-specific patterns consistent with previous literature, supporting both the anatomical fidelity and practical relevance of the results. Two main limitations should be acknowledged. First, our method inherits the constraints of the NextBrain atlas, which was derived from only five cases of older, healthy individuals; consequently, generalization to broader or clinical populations may be limited. Second, although our use of cartoon representations substantially accelerates registration, it may not always achieve the accuracy of probabilistic approaches employed in the original Bayesian algorithm, e.g., in convoluted regions where the probabilistic labels are blurry and yield averaged intensities that may not faithfully represent the underlying tissue. Overall, our new method offers a powerful and accessible tool for fast, high-resolution brain segmentation. By combining detailed histological priors with a lightweight implementation, it bridges the gap between expert-curated atlases and real-world neuroimaging needs. Its speed, simplicity, and robustness make it especially well-suited for large-scale studies at the subregion level. Crucially, by removing the need for specialized software or manual anatomical expertise, this freely available tool democratizes access to fine-grained brain segmentation, enabling a wider community of researchers to explore the brain at unprecedented spatial detail. Data and Code Availability A ready-to-use tool is available as part of FreeSurfer: https://surfer.nmr.mgh.harvard.edu/fswiki/HistoAtlasSegmentation . As with the rest of FreeSurfer, the source code is available on GitHub: https://github.com/freesurfer/freesurfer/tree/dev/mri histo util. AuthorContributions Conceptualization: OP, YB, JEI Data curation: AC, MM, ER, JA, SC, EB, BB, AA, LZ, DLT, DK, MB Formal analysis: AC, LZ, OP, JEI Funding acquisition: OP, JEI Investigation: AC, MM, ER, LP, RA, JA, SC, EB, LZ, YB, OP, JEI Methodology: AC, MM, OP, YB, JLH, ZJ, JEI Project administration: ER, JLH, CS, ZJ, OP, JEI Resources: DLT, JLH, CS, ZJ Software: AC, MM, BB, AA, OP, YB, PS, JH, JN, RD, JEI Supervision: ER, LP, DK, MB, JLH, CS, ZJ, JEI Validation: AC, JEI Visualisation: OP, AC, PS, JH, JEI Writing – original draft: OP, JEI Writing – review and editing: all authors. Funding This research was primarily funded by the NIH (1RF1MH123195, 1R01AG070988, 1UM1MH130981, 1RF1AG080371, 1R21NS138995). Further support was provided by the European Research Council (Starting Grant 677697, project “BUNGEE-TOOLS”). MB was supported by a Fellowship award from the Alzheimer’s Society, UK (AS-JF-19a-004-517) and a grant from the Alzheimer’s Research UK (ARUK-PPG2023B-013). OP was supported by a grant from the Lundbeck foundation (R360–2021–39). MM was supported by the Italian National Institute of Health with a Starting Grant, and by the Wellcome Trust through a Sir Henry Wellcome Fellowship (213722/Z/18/Z). DLT was funded by the National Institute for Health and Care Research University College London Hospitals Biomedical Research Centre (NIHR UCLH BRC) Declaration of Competing Interests The authors have no relevant financial or non-financial interests to disclose. Supplementary Material Heuristic intensity estimates for creating a synthetic MRI scan We first compute the median intensities using the coarse segmentation from BrainFM for the white matter ( µ wm ), gray matter ( µ gm ), cerebellar white matter ( µ cwm ), cerebellar gray matter ( µ cgm ), caudate ( µ ca ), putamen ( µ pu ), and pallidum ( µ pa ). Next the average value and difference between white and gray matter medians is computed: The substructure intensities where no direct label correspondence between the coarse BrainFM segmentation and NextBrain exists are created from µ mid and δ as follows: Ex vivo comparison against the manual labeling on ( Edlow et al., 2019 ) View this table: View inline View popup Table S1: The volumes and Dice scores across all ROIs on the ex vivo sample from Edlow et al., 2019 . Intra-class correlation analysis on the MIRIAD data set View this table: View inline View popup Table S2: The volumes and intra-class correlation coefficients (ICC) over all the ROIs in the NextBrain atlas for the MIRIAD data set. The ROIs are listed in from largest to smallest. Acknowledgements The authors thank Rohit Jena for his prompt and helpful responses to our inquiries regarding his FireANTs method and software. Funder Information Declared National Institutes of Health, https://ror.org/01cwqze88 , 1RF1MH123195 , 1R01AG070988 , 1UM1MH130981 , 1RF1AG080371 , 1R21NS138995 European Research Council, https://ror.org/0472cxd90 , 677697 Alzheimer’s Society , AS-JF-19a-004- 517 AlzheimerM-BM-^Rs Research UK , ARUK-PPG2023B-013 Lundbeck Foundation, https://ror.org/03hz8wd80 , R360M-BM-^V2021M-BM-^V39 Italian National Institute of Health Wellcome Trust, https://ror.org/029chgv08 , 213722/Z/18/Z National Institute for Health and Care Research University College London Hospitals Biomedical Research Centre Footnotes https://surfer.nmr.mgh.harvard.edu/fswiki/HistoAtlasSegmentation References ↵ Ashburner , J. , & Friston , K. J. ( 2005 ). Unified segmentation . neuroimage , 26 ( 3 ), 839 – 851 . OpenUrl CrossRef PubMed Web of Science ↵ Avants , B. B. , Epstein , C. L. , Grossman , M. , & Gee , J. C. ( 2008 ). Symmetric diffeomorphic image registration with crosscorrelation: Evaluating automated labeling of elderly and neurodegenerative brain . Medical image analysis , 12 ( 1 ), 26 – 41 . OpenUrl CrossRef PubMed Web of Science ↵ Bethlehem , R. A. , Seidlitz , J. , White , S. R. , Vogel , J. W. , Anderson , K. M. , Adamson , C. , Adler , S. , Alexopoulos , G. S. , Anagnostou , E. , Areces-Gonzalez , A. , et al. ( 2022 ). Brain charts for the human lifespan . Nature , 604 ( 7906 ), 525 – 533 . OpenUrl CrossRef PubMed ↵ Billot , B. , Greve , D. N. , Puonti , O. , Thielscher , A. , Van Leemput , K. , Fischl , B. , Dalca , A. V. , Iglesias , J. E. , et al. ( 2023 ). SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining . Medical image analysis , 86 , 102789 . OpenUrl CrossRef PubMed ↵ Casamitjana , A. , Mancini , M. , Robinson , E. , Peter , L. , Annunziata , R. , Althonayan , J. , Crampsie , S. , Blackburn , E. , Billot , B. , Atzeni , A. , et al. ( 2025 ). A probabilistic histological atlas of the human brain for mri segmentation . Nature (conditionally accepted) . ↵ Cox , R. W. ( 1996 ). AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages . Computers and Biomedical research , 29 ( 3 ), 162 – 173 . OpenUrl CrossRef PubMed Web of Science ↵ Cuingnet , R. , Gerardin , E. , Tessieras , J. , Auzias , G. , Lehéricy , S. , Habert , M.-O. , Chupin , M. , Benali , H. , Colliot , O. , Initiative , A. D. N. , et al. ( 2011 ). Automatic classification of patients with Alzheimer’s disease from structural MRI: A comparison of ten methods using the ADNI database . neuroimage , 56 ( 2 ), 766 – 781 . OpenUrl CrossRef PubMed Web of Science ↵ Dempster , A. P. , Laird , N. M. , & Rubin , D. B. ( 1977 ). Maximum likelihood from incomplete data via the EM algorithm . Journal of the royal statistical society: series B (methodological) , 39 ( 1 ), 1 – 22 . OpenUrl CrossRef Web of Science ↵ Dickerson , B. , Stoub , T. , Shah , R. , Sperling , R. , Killiany , R. , Albert , M. , Hyman , B. , Blacker , D. , & Detoledo-Morrell , L. ( 2011 ). Alzheimer-signature mri biomarker predicts ad dementia in cognitively normal adults . Neurology , 76 ( 16 ), 1395 – 1402 . OpenUrl CrossRef PubMed ↵ Dufumier , B. , Grigis , A. , Victor , J. , Ambroise , C. , Frouin , V. , & Duchesnay , E. ( 2022 ). OpenBhB: A large-scale multi-site brain MRI data-set for age prediction and debiasing . NeuroImage , 263 , 119637 . OpenUrl PubMed ↵ Edlow , B. L. , Mareyam , A. , Horn , A. , Polimeni , J. R. , Witzel , T. , Tisdall , M. D. , Augustinack , J. C. , Stockmann , J. P. , Diamond , B. R. , Stevens , A. , et al. ( 2019 ). 7 Tesla MRI of the ex vivo human brain at 100 micron resolution . Scientific data , 6 ( 1 ), 244 . OpenUrl PubMed ↵ Fischl , B. ( 2012 ). FreeSurfer . Neuroimage , 62 ( 2 ), 774 – 781 . OpenUrl CrossRef PubMed Web of Science ↵ Fischl , B. , Salat , D. H. , Busa , E. , Albert , M. , Dieterich , M. , Haselgrove , C. , Van Der Kouwe , A. , Killiany , R. , Kennedy , D. , Klaveness , S. , et al. ( 2002 ). Whole brain segmentation: Automated labeling of neuroanatomical structures in the human brain . Neuron , 33 ( 3 ), 341 – 355 . OpenUrl CrossRef PubMed Web of Science ↵ Frisoni , G. B. , Fox , N. C. , Jack Jr , C. R. , Scheltens , P. , & Thompson , P. M. ( 2010 ). The clinical use of structural MRI in Alzheimer disease . Nature reviews neurology , 6 ( 2 ), 67 – 77 . OpenUrl PubMed ↵ Friston , K. J. , Holmes , A. P. , Worsley , K. J. , Poline , J.-P. , Frith , C. D. , & Frackowiak , R. S. ( 1994 ). Statistical parametric maps in functional imaging: A general linear approach . Human brain mapping , 2 ( 4 ), 189 – 210 . OpenUrl CrossRef ↵ Giedd , J. N. , Blumenthal , J. , Jeffries , N. O. , Castellanos , F. X. , Liu , H. , Zijdenbos , A. , Paus , T. , Evans , A. C. , & Rapoport , J. L. ( 1999 ). Brain development during childhood and adolescence: A longitudinal MRI study . Nature neuroscience , 2 ( 10 ), 861 – 863 . OpenUrl CrossRef PubMed Web of Science ↵ Gogtay , N. , Giedd , J. N. , Lusk , L. , Hayashi , K. M. , Greenstein , D. , Vaituzis , A. C. , Nugent III , T. F. , Herman , D. H. , Clasen , L. S. , Toga , A. W. , et al. ( 2004 ). Dynamic mapping of human cortical development during childhood through early adulthood . Proceedings of the national academy of sciences , 101 ( 21 ), 8174 – 8179 . OpenUrl Abstract / FREE Full Text ↵ Gopinath , K. , Hu , X. , Hoffmann , M. , Puonti , O. , & Iglesias , J. E. ( 2024 ). Registration by regression (RbR): A framework for interpretable and flexible atlas registration . International Workshop on Biomedical Image Registration , 205 – 215 . ↵ Greve , D. N. , Billot , B. , Cordero , D. , Hoopes , A. , Hoffmann , M. , Dalca , A. V. , Fischl , B. , Iglesias , J. E. , & Augustinack , J. C. ( 2021 ). A deep learning toolbox for automatic segmentation of subcortical limbic structures from MRI images . Neuroimage , 244 , 118610 . OpenUrl PubMed ↵ Guan , H. , & Liu , M. ( 2021 ). Domain adaptation for medical image analysis: A survey . IEEE Transactions on Biomedical Engineering , 69 ( 3 ), 1173 – 1185 . OpenUrl ↵ Henschel , L. , Conjeti , S. , Estrada , S. , Diers , K. , Fischl , B. , & Reuter , M. ( 2020 ). Fastsurfer-a fast and accurate deep learning based neuroimaging pipeline . NeuroImage , 219 , 117012 . OpenUrl CrossRef PubMed ↵ Hu , X. , Gopinath , K. , Liu , P. , Hoffmann , M. , Van Leemput , K. , Puonti , O. , & Iglesias , J. E. ( 2025 ). Hierarchical uncertainty estimation for learning-based registration in neuroimaging . International Conference on Learning Representations . ↵ Iglesias , J. E. , Augustinack , J. C. , Nguyen , K. , Player , C. M. , Player , A. , Wright , M. , Roy , N. , Frosch , M. P. , McKee , A. C. , Wald , L. L. , et al. ( 2015 ). A computational atlas of the hippocampal formation using ex vivo, ultra-high resolution MRI: Application to adaptive segmentation of in vivo mri . Neuroimage , 115 , 117 – 137 . OpenUrl CrossRef PubMed ↵ Iglesias , J. E. , Insausti , R. , Lerma-Usabiaga , G. , Bocchetta , M. , Van Leemput , K. , Greve , D. N. , Van der Kouwe , A. , Fischl , B. , Caballero-Gaudes , C. , Paz-Alonso , P. M. , et al. ( 2018 ). A probabilistic atlas of the human thalamic nuclei combining ex vivo MRI and histology . Neuroimage , 183 , 314 – 326 . OpenUrl CrossRef PubMed ↵ Isensee , F. , Jaeger , P. F. , Kohl , S. A. , Petersen , J. , & Maier-Hein , K. H. ( 2021 ). nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation . Nature methods , 18 ( 2 ), 203 – 211 . OpenUrl PubMed Jack Jr , C. R. , Petersen , R. C. , Xu , Y. C. , O’Brien , P. C. , Smith , G. E. , Ivnik , R. J. , Boeve , B. F. , Waring , S. C. , Tangalos , E. G. , & Kokmen , E. ( 1999 ). Prediction of AD with MRI-based hippocampal volume in mild cognitive impairment . Neurology , 52 ( 7 ), 1397 – 1397 . OpenUrl CrossRef PubMed ↵ Jena , R. , Chaudhari , P. , & Gee , J. C. ( 2024 ). Fireants: Adaptive riemannian optimization for multi-scale diffeomorphic matching . arXiv preprint arXiv: 2404.01249 . ↵ Kingma , D. P. , & Ba , J. ( 2014 ). Adam: A method for stochastic optimization . arXiv preprint arXiv: 1412.6980 . ↵ Liu , P. , Puonti , O. , Hu , X. , Gopinath , K. , Sorby-Adams , A. , Alexander , D. C. , Kimberly , W. T. , & Iglesias , J. E. ( 2025 ). A modality-agnostic multi-task foundation model for human brain imaging . arXiv preprint arXiv: 2509.00549 . ↵ Malone , I. B. , Cash , D. , Ridgway , G. R. , MacManus , D. G. , Ourselin , S. , Fox , N. C. , & Schott , J. M. ( 2013 ). Miriad—public release of a multiple time point alzheimer’s mr imaging dataset . NeuroImage , 70 , 33 – 36 . OpenUrl CrossRef PubMed ↵ Nocedal , J. , & Wright , S. J. ( 2006 ). Numerical optimization . Springer . ↵ Pereira , S. , Pinto , A. , Alves , V. , & Silva , C. A. ( 2016 ). Brain tumor segmentation using convolutional neural networks in MRI images . IEEE transactions on medical imaging , 35 ( 5 ), 1240 – 1251 . OpenUrl CrossRef PubMed ↵ Pohl , K. M. , Fisher , J. , Grimson , W. E. L. , Kikinis , R. , & Wells , W. M. ( 2006 ). A Bayesian model for joint segmentation and registration . NeuroImage , 31 ( 1 ), 228 – 239 . OpenUrl PubMed ↵ Puonti , O. , Iglesias , J. E. , & Van Leemput , K. ( 2016 ). Fast and sequence-adaptive whole-brain segmentation using parametric bayesian modeling . NeuroImage , 143 , 235 – 249 . OpenUrl CrossRef PubMed ↵ Ratiu , P. , Hillen , B. , Glaser , J. , & Jenkins , D. P. ( 2003 ). Visible Human 2.0 – the next generation . In Medicine meets virtual reality 11 (pp. 275 – 281 ). IOS Press . OpenUrl ↵ Ronneberger , O. , Fischer , P. , & Brox , T. ( 2015 ). U-net: Convolutional networks for biomedical image segmentation . International Conference on Medical image computing and computer-assisted intervention , 234 – 241 . ↵ Salehi , S. S. M. , Erdogmus , D. , & Gholipour , A. ( 2017 ). Tversky loss function for image segmentation using 3D fully convolutional deep networks . International workshop on machine learning in medical imaging , 379 – 387 . ↵ Saygin , Z. M. , Kliemann , D. , Iglesias , J. E. , van der Kouwe , A. J. , Boyd , E. , Reuter , M. , Stevens , A. , Van Leemput , K. , McKee , A. , Frosch , M. P. , et al. ( 2017 ). High-resolution magnetic resonance imaging reveals nuclei of the human amygdala: Manual segmentation to automatic atlas . Neuroimage , 155 , 370 – 382 . OpenUrl CrossRef PubMed ↵ Smith , S. M. , Jenkinson , M. , Woolrich , M. W. , Beckmann , C. F. , Behrens , T. E. , Johansen-Berg , H. , Bannister , P. R. , De Luca , M. , Drobnjak , I. , Flitney , D. E. , et al. ( 2004 ). Advances in functional and structural MR image analysis and implementation as FSL . Neuroimage , 23 , S208 – S219 . OpenUrl CrossRef PubMed Web of Science ↵ Van Leemput , K. , Maes , F. , Vandermeulen , D. , & Suetens , P. ( 2002 ). Automated model-based tissue classification of MR images of the brain . IEEE transactions on medical imaging , 18 ( 10 ), 897 – 908 . OpenUrl ↵ Walsh , C. L. , Tafforeau , P. , Wagner , W. , Jafree , D. , Bellier , A. , Werlein , C. , Kühnel , M. , Boller , E. , Walker-Samuel , S. , Robertus , J. , et al. ( 2021 ). Imaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography . Nature methods , 18 ( 12 ), 1532 – 1541 . OpenUrl PubMed ↵ Wang , M. , & Deng , W. ( 2018 ). Deep visual domain adaptation: A survey . Neurocomputing , 312 , 135 – 153 . OpenUrl CrossRef ↵ Wells , W. M. , Grimson , W. E. L. , Kikinis , R. , & Jolesz , F. A. ( 1996 ). Adaptive segmentation of MRI data . IEEE transactions on medical imaging , 15 ( 4 ), 429 – 442 . OpenUrl CrossRef PubMed Web of Science ↵ Zhang , Y. , Brady , M. , & Smith , S. ( 2002 ). Segmentation of brain MR images through a hidden markov random field model and the expectation-maximization algorithm . IEEE transactions on medical imaging , 20 ( 1 ), 45 – 57 . OpenUrl CrossRef View the discussion thread. Back to top Previous Next Posted September 25, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Fast segmentation with the NextBrain histological atlas Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Fast segmentation with the NextBrain histological atlas Oula Puonti , Jackson Nolan , Robert Dicamillo , Yael Balbastre , Adria Casamitjana , Matteo Mancini , Eleanor Robinson , Loic Peter , Roberto Annunziata , Juri Althonayan , Shauna Crampsie , Emily Blackburn , Benjamin Billot , Alessia Atzeni , Peter Schmidt , James Hughes , Jean Augustinack , Brian Edlow , Lilla Zöllei , David L Thomas , Dorit Kliemann , Martina Bocchetta , Catherine Strand , Janice Holton , Zane Jaunmuktane , Juan Eugenio Iglesias bioRxiv 2025.09.22.673638; doi: https://doi.org/10.1101/2025.09.22.673638 Share This Article: Copy Citation Tools Fast segmentation with the NextBrain histological atlas Oula Puonti , Jackson Nolan , Robert Dicamillo , Yael Balbastre , Adria Casamitjana , Matteo Mancini , Eleanor Robinson , Loic Peter , Roberto Annunziata , Juri Althonayan , Shauna Crampsie , Emily Blackburn , Benjamin Billot , Alessia Atzeni , Peter Schmidt , James Hughes , Jean Augustinack , Brian Edlow , Lilla Zöllei , David L Thomas , Dorit Kliemann , Martina Bocchetta , Catherine Strand , Janice Holton , Zane Jaunmuktane , Juan Eugenio Iglesias bioRxiv 2025.09.22.673638; doi: https://doi.org/10.1101/2025.09.22.673638 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7622) Biochemistry (17650) Bioengineering (13871) Bioinformatics (41880) Biophysics (21424) Cancer Biology (18566) Cell Biology (25461) Clinical Trials (138) Developmental Biology (13365) Ecology (19866) Epidemiology (2067) Evolutionary Biology (24290) Genetics (15590) Genomics (22475) Immunology (17713) Microbiology (40328) Molecular Biology (17148) Neuroscience (88473) Paleontology (666) Pathology (2827) Pharmacology and Toxicology (4816) Physiology (7635) Plant Biology (15114) Scientific Communication and Education (2044) Synthetic Biology (4286) Systems Biology (9815) Zoology (2268)