TLDR: NEUBORN is a novel deep learning framework that uses biomechanical constraints to accurately model individual brain growth trajectories from longitudinal MRI scans. It improves upon existing methods by generating more biologically plausible deformations, reducing anatomical errors, and faithfully tracking population-level growth trends, offering a powerful tool for early detection of neurodevelopmental disorders.
Understanding how the human brain develops, particularly the intricate folding patterns of the cerebral cortex, is crucial for identifying early signs of neurodevelopmental disorders. Current methods for modeling brain development often struggle to capture the fine details of individual brain anatomy because they rely on averaging data across many individuals. This can lead to a loss of precision when trying to detect subtle abnormalities unique to a single person.
The process of cortical folding is complex. While the primary folds are largely determined by genetics, the formation of secondary and tertiary folds is significantly influenced by biomechanical forces. Even minor disturbances in geometry or mechanical properties, such as differences in cortical thickness or growth rates, can drastically alter the final brain shape. This sensitivity explains why even identical twins can have distinct cortical folding patterns.
Errors during cortical folding can result in Malformations of Cortical Development (MCDs), which are linked to various neurodevelopmental disorders like intellectual disability, autism spectrum conditions, cerebral palsy, and epilepsy. Existing approaches for detecting early biomarkers often compare brain features within a population-average space. However, the high variability in individual cortical folding makes it difficult to perfectly align all brains, leading to spatial blurring and reduced sensitivity in detecting subtle, subject-specific biomarkers.
Introducing NEUBORN
A new framework called NEUBORN (Neurodevelopmental Evolution framework Using BiOmechanical RemodelliNg) has been developed to address these challenges. This novel approach learns individual brain growth trajectories by using biomechanically constrained, longitudinal, diffeomorphic image registration. It is implemented through a hierarchical network architecture, designed to handle the large deformations that occur during brain growth.
NEUBORN was trained using neonatal MRI data from the Developing Human Connectome Project. The method significantly improves the biological realism of the brain transformations it generates. It produces growth trajectories that align better with population-level trends while also creating smoother transformations with fewer “negative Jacobians,” which are indicators of anatomically impossible tissue folding or tearing. The resulting subject-specific deformations offer interpretable and biologically sound maps of brain development.
How NEUBORN Works
The framework uses a deep learning approach, specifically a neural network, to map a moving image (an earlier brain scan) into alignment with a reference image (a later brain scan). The network is trained to minimize two main factors: the difference in appearance between the warped image and the reference, and a biomechanical loss that ensures the deformations are physiologically plausible. This biomechanical loss is based on a Neo-Hookean model, which considers the mechanical properties of different brain tissues like grey matter, white matter, and cerebrospinal fluid.
To handle the significant changes in brain shape during development, NEUBORN employs a hierarchical architecture with two U-Nets operating at different resolutions. A low-resolution network first learns large-scale deformations, and its output is then refined by a high-resolution network that captures finer details. This multi-resolution approach allows the model to accurately track complex cortical growth.
Key Findings
In evaluations, NEUBORN demonstrated alignment accuracy comparable to state-of-the-art registration techniques like VoxelMorph. However, a crucial advantage of NEUBORN is its dramatic reduction in anatomically implausible transformations, producing approximately four orders of magnitude fewer negative Jacobians than VoxelMorph. This means NEUBORN’s transformations are much more biologically realistic.
Furthermore, NEUBORN’s predictions of brain volume changes (specifically cortical grey matter and white matter) closely followed observed population-level growth trends. When comparing simulated follow-up volumes to actual ground truth volumes, NEUBORN showed minimal differences, indicating its ability to accurately model individual growth trajectories. In contrast, other methods like VoxelMorph tended to overestimate cortical volume and showed larger deviations from true individual trajectories.
Also Read:
- Improved Stochastic Function Learning with Neural Bridge Processes
- PRISM: A New Foundation Model for Comprehensive MRI Analysis
Implications and Future Directions
The NEUBORN framework represents a significant step forward in longitudinal image registration. By integrating a diffeomorphic registration strategy with a physically grounded biomechanical loss, it generates anatomically plausible deformations that accurately preserve complex subject-specific cortical growth trajectories. This capability is particularly valuable for identifying subtle deviations from typical development, which could be early indicators of neurological disorders.
Future work will explore leveraging NEUBORN’s biomechanical interpretability to model intermediate developmental states between scans and to enable forward projections of future brain changes at the patient level. For more in-depth technical details, you can refer to the full research paper available at arXiv.
