Abstract
Effective representation of brain anatomical architecture is fundamental in understanding brain regularity and variability. Despite numerous efforts, it is still difficult to infer reliable anatomical correspondence at a finer scale, given the tremendous individual variability in cortical folding patterns. It is even more challenging to disentangle common and individual patterns when comparing brains at different neuro-developmental stages. In this work, we developed a novel learning-based few-shot embedding framework to encode the cortical folding patterns into a latent space represented by a group of anatomically meaningful embedding vectors. Specifically, we adopted the 3-hinge (3HG) network as the substrate and designed an autoencoder-based embedding framework to learn a common embedding vector for each 3HG's multi-hop feature. Each 3HG can be represented as a combination of these feature embeddings via a set of individual-specific coefficients to characterize individualized anatomical information. That is, the regularity of folding patterns is encoded into the embeddings, while the individual variations are preserved by the multi-hop combination coefficients. To effectively learn the embeddings for populations with very limited samples, few-shot learning was adopted. We applied our method on adult HCP and pediatric datasets with 1,000+ brains (from 34 gestational weeks to young adult). Our experimental results show that: 1) the learned embedding vectors can quantitatively encode the commonality and individuality of cortical folding patterns; 2) with the embeddings we can robustly infer the complicated many-to-many anatomical correspondences among different brains; and 3) our model can be successfully transferred to new populations with very limited training samples.
Introduction
Understanding the human brain's structural organization requires precise mapping of cortical folding patterns—the complex ridges (gyri) and grooves (sulci) that define the brain's surface. However, two major hurdles impede progress: the correspondence problem and data scarcity. Because every brain folds uniquely, establishing exact anatomical matches between individuals, especially across different developmental stages, is notoriously difficult. Furthermore, while large-scale datasets like the Human Connectome Project (HCP) provide ample data for adults, high-quality scans for infants at critical developmental windows (e.g., 34 gestational weeks) remain extremely rare.
Traditional methods often rely on local geometric features like curvature or global averages, which frequently fail to capture the fine-grained "regularity" (shared features) and "variability" (unique features) simultaneously. This research introduces a few-shot learning framework that maps these complex patterns into a shared mathematical space, enabling robust comparison across the entire human lifespan.
Methodology
The proposed framework utilizes a novel architecture designed to disentangle shared anatomical templates from individual deviations through four key components:
- 3-Hinge (3HG) Networks: The model focuses on "3-hinges"—junctions where three gyri meet. These serve as stable anatomical landmarks across individuals.
- Multi-Hop Feature Generation: To provide spatial context, the model incorporates information from a landmark's neighbors (1-hop) and their subsequent neighbors (2-hop), represented as a graph based on the Destrieux Atlas.
- Self-Supervised Autoencoder: An autoencoder compresses these multi-hop features into a latent embedding space. Anatomical regularity is captured in shared embedding vectors, while variability is preserved in unique coefficients used to reconstruct an individual's specific brain architecture.
- Sequential Transfer Learning: To overcome data scarcity, the model is first trained on over 1,000 adult brains. This "knowledge" is then sequentially transferred to 2-year-olds, 6-month-olds, and finally 34-week-old fetuses. This approach leverages the biological reality that brain development is a continuous process where adult folding patterns provide a structural baseline for infant folds.
Datasets
The study utilized a diverse range of neuroimaging data:
- Adults: 1,064 subjects from the Human Connectome Project (HCP) S1200 release.
- Pediatric: Data sourced from the National Institute of Mental Health Data Archive (NDA) and the developing Human Connectome Project (dHCP), including 12 two-year-olds, 10 six-month-olds, and 8 subjects at 34 gestational weeks.
Experimental Evaluation
The model was evaluated against "from-scratch" training benchmarks to determine the efficacy of the few-shot transfer approach. The results demonstrated significant improvements in both accuracy and efficiency.
Performance and Convergence
When applied to the 34-week-old group using only two training subjects, the few-shot model achieved a Mean Square Error (MSE) significantly lower than models trained from scratch. Notably, the few-shot models converged in fewer than 30 epochs, whereas from-scratch models failed to reach comparable accuracy even after 300+ epochs.
Structural Similarity and Correspondence
The learned embeddings exhibited high Structural Similarity Index (SSIM) scores with ground-truth anatomical connections—reaching up to 0.84 for 1-hop features. In contrast, from-scratch models performed poorly on infant data, with SSIM scores dropping as low as 0.07. Qualitative analysis confirmed that the model could successfully identify corresponding anatomical points across different individuals and age groups with high confidence (cosine similarity > 0.9).
Discussion and Future Work
This framework demonstrates that anatomical regularity is preserved across ages, allowing models to leverage adult data to understand fetal brain structure. By disentangling the "template" from individual "deviations," the researchers have enabled more accurate many-to-many anatomical mapping.
While highly effective, the model currently relies on the Destrieux Atlas for initial ROI labeling, which may introduce biases if the atlas does not translate perfectly to fetal brains. Future work will explore the application of these embeddings as early biomarkers for neurodevelopmental disorders such as autism or ADHD, where an individual's variability might deviate significantly from healthy population regularity.
Summary updated