AtlasMorph: Learning conditional deformable templates for brain MRI
Marianne Rakic, Andrew Hoopes, S. Mazdak Abulnaga, Mert R. Sabuncu, John V. Guttag, Adrian V. Dalca, for the Alzheimer’s Disease Neuroimaging Initiative* · 2025
Read AtlasMorph: Learning conditional deformable templates for brain MRI by Marianne Rakic, Andrew Hoopes, S. Mazdak Abulnaga, Mert R. Sabuncu, John V. Guttag, Adrian V. Dalca, for the Alzheimer’s Disease Neuroimaging Initiative* (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-13609.txt.
Introduction
Marianne Rakic and 5 co-authors present AtlasMorph, a learning framework that builds brain MRI templates matched to a chosen population. A template, or atlas, is a prototypical anatomy image used for registration and segmentation across studies. Building one is costly, so researchers often settle for atlases that fit their subjects poorly.
AtlasMorph trains convolutional registration networks to output a template conditioned on traits such as age and sex, with anatomical label maps when segmentations exist. The same network can register new scans to its templates. On a pooled set of 3D brain datasets, conditional labeled templates registered better than unconditional ones and beat other construction methods, giving studies atlases that truly mirror their populations.
Catalog record
| CREATOR | Marianne Rakic, Andrew Hoopes, S. Mazdak Abulnaga, Mert R. Sabuncu, John V. Guttag, Adrian V. Dalca, for the Alzheimer’s Disease Neuroimaging Initiative* |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-13609
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AtlasMorph: Learning conditional deformable templates for brain MRI Marianne Rakic, Andrew Hoopes, S. Mazdak Abulnaga, Mert R. Sabuncu, John V. Guttag, Adrian V. Dalca, for the Alzheimer’s Disease Neuroimaging Initiative* Abstract Deformable templates, or atlases, are images that represent a prototypical anatomy for a population, and are often enhanced with probabilistic anatomical label maps. They are commonly used in medical image analysis for population studies and computational anatomy tasks such as registration and segmentation. Because developing a template is a computationally expensive process, relatively few templates are available. As a result, analysis is often conducted with sub-optimal templates that are not truly representative of the study population, especially when there are large variations within this population. We propose a machine learning framework that uses convolutional registration neural networks to efficiently learn a function that outputs templates conditioned on subject-specific attributes, such as age and sex. We also leverage segmentations, when available, to produce anatomical segmentation maps for the resulting templates. The learned network can also be used to register subject images to the templates. We demonstrate our method on a compilation of 3D brain MRI datasets, and show that it can learn high-quality templates that are representative of populations. We find that annotated conditional templates enable better registration than their unlabeled unconditional counterparts, and outperform other templates construction methods. 0 0 footnotetext: Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf 1 Introduction Anatomical templates are representative medical images that are widely used for population studies and image analysis. In population studies, a template provides a common coordinate system to which the images in the study can be registered. This facilitates constructing statistical models of the population, and analyzing differences between each subject and the common template. Developing a template has historically involved a computationally intensive optimization process, including iterative image registration and domain-specific heuristics 3 ; 26 ; 27 ; 49 ; 58 ; 75 . Because it is labor intensive and computationally expensive, there is often at most one template available per population. This is problematic for some analyses because a single template does not capture the distinctive attributes of sub-populations and can be inadequate at capturing the variability in a large dataset. Attempts to address this problem have typically involved identifying subpopulations, usually along a single attribute such as age, and computing separate templates for each group 75 ; 25 ; 39 ; 54 . This approach relies on arbitrary decisions about the attributes and thresholds used for subdividing the population. Furthermore, since each template is constructed using only images from the subpopulation, it can produce sub-optimal templates – especially for subpopulations with a small number of images. Templates with segmentation, or label map, are widely used by different neuroimaging software such as FreeSurfer 36 , SPM 67 , FSL 80 and BrainSuite 76 . SPM and FreeSurfer use templates involving probabilistic label maps to automatically segment new subject scans 36 . Methods often align the template to the new image and propagate the template label maps, sometimes combined with an image intensity likelihood model 69 . In longitudinal population studies, label maps are crucial to characterize changes in anatomy. However, template label maps are even more challenging to construct than intensity-based templates. There are often many fewer labeled images than unlabeled images for a subpopulation. Consequently, building a useful conditional atlas is more challenging when segmentation is required. In this work, we present AtlasMorph , a learning strategy to generate conditional templates and corresponding anatomical label maps as a function of subject attributes . While we believe the strategy is general, in this paper we demonstrate it only on brain anatomy. Our method can leverage all the data available when constructing templates. There is no need to train a different network on a subset of the data for each combination of attributes. Figure 1 shows examples of templates, generated on demand by sampling our trained age-conditioned template function. Our framework can also adjust when label maps or subject attributes are missing. Figure 1 : Conditional Templates developed with AtlasMorph as a function of age. This paper builds on the methods we introduced in a preliminary conference proceeding 22 . This paper has the following new contributions. • We extend our probabilistic model 22 to characterize an anatomical label map template in addition to an intensity template. To achieve this, we adapt the learning framework to a setting where auxiliary segmentations, when available, can be used during training to build a segmentation template. Label maps are not required for test subjects during inference. • We minimize the centrality of the generated template, i.e., the weighted mean deformation between the (conditional) templates and subjects sharing the same attributes. This addresses a limitation in our previous work 22 , where centrality was improperly modeled for conditional templates. We introduce a new mathematical model, and a data sampling strategy based on continuous attributes, to characterize population centrality. • In a set of experiments on a varied collection of publicly available 3D brain MRI datasets, we show that the templates 1) are representative of the intended population 2) lead to higher registration quality than baseline conditional (age and sex) and unconditional templates, and 3) match better population trends than previous work 22 . 2 Related Work We first discuss related work for registration, which is used both in template building and in downstream applications of templates. We then focus on previous approaches to template construction. 2.1 Registration Image registration produces a dense deformation field that maps one (often called moving) image to another (fixed) image. Registration can be divided into two steps: global affine registration and deformable registration. We work with data that has already been affinely aligned, and focus on deformable transformations. Deformable registration has been widely studied 7 ; 8 ; 11 ; 14 ; 24 ; 37 ; 84 ; 93 ; 91 ; 56 ; 92 ; 78 . Popular classical deformable registration methods include Demons 87 , ANTs 9 , HAMMER 77 , and ELASTIX 50 . These methods solve a new optimization problem for each new image pair, which makes them computationally intensive 11 ; 84 . In contrast, learning-based methods learn a function that for each input image pair outputs the corresponding deformation field. During inference, this provides a considerable speed up over classical methods 12 ; 13 ; 18 ; 23 ; 52 ; 73 ; 81 ; 88 ; 90 ; 55 ; 34 ; 51 ; 43 ; 44 ; 63 . Some learning-based methods are supervised, where “ground truth” deformation fields are precomputed or synthesized 33 ; 19 ; 90 ; 55 . Unsupervised methods learn to predict a dense deformation field without the corresponding ground truth 12 ; 35 ; 28 ; 51 ; 70 ; 59 . Instead, during training they usually employ loss terms to encourage 1) matching between the warped source and the target and 2) regularity