Anatomical templates are widely used in neuroimaging to compare brain structures across individuals. They provide a shared reference that supports image registration, segmentation and population analysis. Traditional template construction relies on computationally intensive processes and usually produces a single template for an entire population. Such templates may not adequately represent anatomical differences across age, sex or disease status. AtlasMorph introduces a machine learning framework that generates deformable brain MRI templates conditioned on subject-specific attributes. By learning templates directly from large imaging datasets, the approach enables more representative anatomical modelling across diverse populations.
Learning Conditional Templates from Population Data
AtlasMorph generates templates as a function of subject attributes rather than dividing populations into fixed subgroups. Conventional approaches often create separate templates for predefined categories, such as age ranges, which can lead to suboptimal results when subgroup data are limited or when attributes vary continuously. AtlasMorph instead learns a single model that can produce templates on demand for specific attribute values, allowing all available data to contribute to template construction.
Deformable templates represent typical anatomy for a population and are frequently combined with probabilistic anatomical label maps. These maps support automated segmentation and longitudinal analysis but require manually annotated images that are often limited. AtlasMorph incorporates segmentation information when available during training to jointly learn intensity templates and anatomical label maps. When segmentation labels are unavailable, the model can still learn intensity templates and later derive label maps through deformation-based alignment.
The framework was evaluated using 10,195 T1-weighted three-dimensional brain MRI scans collected from multiple public datasets, including ADNI, OASIS, ABIDE and IXI. The scans covered subjects aged 12 to 90 years. After preprocessing steps such as affine alignment, skull stripping and anatomical segmentation, the dataset was divided into training, validation and test sets. Analysis focused on 17 anatomical brain structures defined according to the FreeSurfer protocol.
Templates generated across different ages captured expected anatomical variation. Younger-age templates showed greater grey matter representation, while templates representing older subjects displayed enlarged ventricles. These patterns reflect anatomical variability across the lifespan and demonstrate the ability of conditional template generation to model population differences.
Neural Registration and Template Learning
AtlasMorph combines two neural networks: one that generates templates from subject attributes and another that aligns subject images to those templates. The template network produces both intensity images and segmentation probability maps, while the registration network estimates deformation fields that map subject images to the generated template.
The framework uses diffeomorphic deformation fields, which preserve anatomical topology and allow inverse transformations between subject images and templates. Training is guided by a probabilistic model that treats subject images as deformed versions of an attribute-conditioned template. The training objective balances image similarity, segmentation agreement, deformation smoothness and template centrality.
Centrality describes how well a template represents subjects with similar attributes. Earlier approaches measured centrality across the entire population, which could bias templates when anatomical variation depends on attributes such as age. AtlasMorph introduces a conditional centrality formulation that compares deformation fields only among subjects with similar attributes. For continuous variables such as age, kernel-based weighting is used to approximate population averages. This produces templates that better represent subpopulations.
The registration network uses three-dimensional UNet-like architecture, and deformation fields are computed using scaling-and-squaring integration of stationary velocity fields. After training, templates and deformation fields can be generated quickly through forward passes of the neural networks, reducing the computational burden associated with classical iterative template construction.
Must Read: Continual Learning for Medical Image Analysis
Performance In Brain MRI Template Construction
AtlasMorph templates were evaluated using several criteria, including registration accuracy, deformation regularity, template sharpness and the ability to capture population trends. Registration performance was assessed by propagating template segmentation maps to subject images and comparing them with anatomical labels using Dice score and surface distance.
Conditional templates produced by AtlasMorph were sharper and more representative than templates obtained by simple averaging or classical construction methods. Models that jointly learned segmentation maps and intensity templates achieved better performance than models trained without segmentation information. Conditional templates also showed slightly improved results compared with unconditional templates.
Performance varied across age groups, with lower segmentation accuracy observed in younger subjects due to fewer available training examples. Nevertheless, conditional template learning provided consistent results across the population. Deformation fields remained smooth and anatomically plausible, with only minimal numerical irregularities.
AtlasMorph also demonstrated improved ability to capture population anatomical trends. Age-related changes in structures such as ventricles and hippocampi were reflected in the generated templates and closely matched trends estimated from manual segmentations. The conditional centrality formulation produced templates that aligned more closely with population averages than earlier template-learning methods.
Additional experiments showed that template initialisation affects performance. Initialising with the average of multiple subjects produced more stable results than starting from a single subject image. The framework was also extended to include cognitive status attributes, enabling template generation for cognitively normal individuals, mild cognitive impairment and Alzheimer’s disease across different ages.
AtlasMorph enables efficient learning of conditional anatomical templates and segmentation maps from large brain MRI datasets. By conditioning template generation on subject attributes, the framework produces more representative anatomical models than single-template approaches. Joint learning of registration and segmentation improves alignment performance while reducing the need for computationally intensive template-construction methods. Conditional template generation provides a flexible approach for population neuroimaging analysis and supports improved characterisation of anatomical variation in medical imaging research.
Source: Medical Image Analysis
Image Credit: iStock
