When analyzing a collection of biological shapes—whether they are human brains, facial structures, or dental arches—researchers must establish a common coordinate system to compare them. The traditional way to do this is through pairwise registration, where a single subject is chosen as a "gold standard" reference template, and every other subject is warped to match it.
While straightforward, this approach suffers from a fundamental flaw known as template-induced bias (or reference bias). Because the optimization process forces every shape to conform to a single geometry, the final population model becomes systematically skewed toward the unique idiosyncrasies, topology, and artifacts of the chosen template. If the template happens to be an anatomical outlier, the precision of the entire study is compromised.
To overcome this, a major research direction focuses on groupwise registration and template-free alignment. Instead of selecting an arbitrary anchor, these methods align the entire population simultaneously into a neutral, dynamically computed common space, ensuring that no single subject dominates the group's representation.
The Geometry of Reference Bias
Why exactly does a fixed template introduce bias? In pairwise registration, the similarity metric is asymmetric: it measures how well a moving image transforms into a static target. As noted by Dalal et al. in Multiple Cortical Surface Correspondence Using Pairwise Shape Similarity, fixed-template matching fails to accurately map individuals whose structures deviate significantly from the reference. This constraint distorts natural biological variations, a problem highlighted by Namburete in Normative spatiotemporal fetal brain maturation with satisfactory development at 2 years.
This distortion has real-world clinical consequences. For instance, Shan et al. demonstrated in Anthropometric accuracy of three-dimensional average faces compared to conventional facial measurements that using an individual facial scan as a registration template introduces systematic errors that skew downstream clinical measurements.
Pairwise (Biased): Subject 1 (Template) <--- Subject 2, Subject 3, Subject 4
Groupwise (Unbiased): Subject 1, 2, 3, 4 ---> [ Latent / Evolving Mean Space ]
Groupwise registration fixes this asymmetry by treating all subjects equally. By optimizing the alignment of all images simultaneously toward a shared, evolving center, the mathematical gradients pull toward the true arithmetic mean of the population rather than an arbitrary individual anchor.
Mathematical Formulations for Joint Alignment
Simultaneously aligning dozens or hundreds of high-dimensional images introduces a massive computational challenge. Researchers have developed distinct mathematical frameworks to make this joint optimization scalable and robust.
Information-Theoretic Metrics
One approach is to minimize the joint entropy across the image stack. Bhatia et al. proposed a scalable metric called Average Normalised Mutual Information (ANMI) in Similarity Metrics for Groupwise Non-rigid Registration, which allows researchers to align large groups of images into an unbiased atlas without the computational explosion typical of multi-dimensional entropy calculations. For unlabelled point clouds, Hasanbelliu leveraged Rényi’s second-order entropy in Group-Wise Point-Set Registration Based on Rényi's Second Order Entropy to create a template-free framework that treats point sets as probability distributions, achieving objective alignment without requiring a target shape.
Iterative Bootstrapping
When an explicit population average is needed but no unbiased starting point exists, researchers use iterative feedback loops. In Statistical Atlases of Bone Anatomy: Construction, Iterative Improvement and Validation, Chintalapani et al. introduced a bootstrapping framework that uses an initial, rough statistical model to guide subsequent registrations. As the registrations improve, the central model is updated, progressively washing away the bias of the initial shape. Modern deep learning architectures, such as the one proposed by Iyer et al. in Mesh2SSM: From Surface Meshes to Statistical Shape Models of Anatomy, automate this process via data-driven template feedback loops to capture complex, non-linear variations directly from raw surface meshes.
Aligning in Functional and Latent Spaces
A more radical strategy to eliminate template bias is to abandon physical coordinate templates altogether, shifting the alignment problem into abstract functional or latent spaces.
Instead of matching spatial coordinates point-by-point, the authors of Functional Faces: Groupwise Dense Correspondence Using Functional Maps (Zhang et al.) operated in the functional map domain. By matching real-valued functions across surfaces (such as texture and geometric curvature) through a "virtual universe" of points, they achieved guaranteed cycle consistency across entire datasets without an explicit template mesh.
Physical Space: Mesh A --------(Template Bias)--------> Mesh B
Functional Space: Mesh A ---> [ Abstract Functional Domain ] <--- Mesh B
Similarly, Thual et al. utilized Optimal Transport in Aligning individual brains with fused unbalanced Gromov Wasserstein to align functional brain maps. By calculating the Wasserstein distance between individuals based on both functional similarity and anatomical geometry, their method bypasses intermediate average brain templates entirely, preserving fine-grained, subject-specific features that traditional templates obscure.
In the domain of deep generative modeling, Starck introduced CINeMA: Conditional Implicit Neural Multi-Modal Atlas for a Spatio-Temporal Representation of the Perinatal Brain, which leverages Implicit Neural Representations (INRs). By encoding anatomical variations within a continuous latent space, this framework eliminates the traditional requirement of pre-aligning individual subject images to a fixed template before modeling.
Eliminating Bias in Time-Series and Clinical Pipelines
The problem of reference bias compounds when data features a temporal or longitudinal dimension. In studies tracking disease progression or fetal growth, selecting the first time point or a baseline scan as the reference template biases the analysis against detecting subsequent structural changes.
To resolve this in longitudinal sequences, Wu proposed an implicit template framework in Registration of Longitudinal Image Sequences with Implicit Template and Spatial-Temporal Heuristics. This method simultaneously aligns images from multiple subjects across multiple time points into a "hidden" common space, preventing any single scan from distorting the trajectory of anatomical changes. When tracking highly unpredictable, large-scale deformations—such as fetal movement in 4D MRI sequences—Liao et al. demonstrated in Temporal Registration in In-Utero Volumetric MRI Time Series that modeling the sequence via a Hidden Markov Model allows the registration to adapt dynamically to frame-to-frame changes, avoiding the constraints and biases of a static reference frame.
Ultimately, moving from single-template matching to groupwise, template-free architectures ensures that population-level analyses reflect true biological variance rather than the mathematical artifacts of an arbitrarily chosen reference.