International Transfer of Stochastic Cortical Self-Reconstruction

This paper investigates the international transferability of Stochastic Cortical Self-Reconstruction (SCSR), a neural network-based normative model for personalized mapping of gray matter atrophy on cortical surfaces, originally trained on UK Biobank (UKB) data, to an independent Chinese population dataset.

Background and motivation. Cortical thickness is a sensitive biomarker for distinguishing dementia types and predicting progression from mild cognitive impairment (MCI) to Alzheimer’s disease (AD), but identifying individual-level changes is difficult due to inter-subject variability and scanner/acquisition effects. Classical normative models estimate expected brain measurements as a function of covariates like age and sex, yet they remain inherently limited to the considered covariates, produce references not tailored to an individual's unique brain morphology, and do not scale easily to thousands of cortical thickness measurements. SCSR overcomes these limits: instead of mapping demographics to an expected phenotype, it repeatedly reconstructs an individual’s full cortical morphology from a stochastically sampled subset of measurements, using a neural network trained exclusively on healthy individuals. A personalized reference distribution is obtained via repeated sampling, from which vertex-wise Z-scores are computed.

Data. UK Biobank imaging data (25,338 healthy training subjects, 6,334 validation) covers ages 45–82. The Chinese population dataset spans ages 4–86 and includes 640 healthy scans for training/finetuning, 160 for validation, a lifespan test set of 139 subjects stratified into 10-year age brackets, and for atrophy detection: 60 subjects each per diagnostic group (AD, MCI, cognitively normal controls aged 46–85). All scans were processed with FreeSurfer, yielding 10,242 vertex-wise thickness values per hemisphere.

Methods. SCSR samples a fraction s=20% of vertices as predictors, reconstructs the remaining vertices, and is trained to minimize reconstruction error of masked vertices. At test time, m=100 stochastic repetitions produce a tensor, and the subject-specific reference R is the vertex-wise reconstruction centile q=0.95; Z-scores are computed as (Y−R)/σ. Two reconstruction backbones are compared: a multilayer perceptron (MLP) with 20,055,608 parameters (roughly twelve times larger than the SUNet, but does not encode any spatial inductive bias) and a Spherical UNet (SUNet) with 1,669,217 parameters performing spherical convolutions on the icosahedral mesh. Four adaptation strategies are evaluated for each backbone: direct application of the UKB-trained model, fine-tuning on Chinese data, training from scratch on Chinese data, and joint training on pooled UKB + Chinese data, yielding eight configurations. Evaluation metrics are the mean absolute reconstruction error (MAE) between observed thickness and SCSR reference, and the area under the ROC curve (AUC) for pairwise diagnostic discrimination (CN vs. MCI, CN vs. AD, MCI vs. AD) computed on mean Z-scores within an AD ROI (entorhinal, inferior temporal, middle temporal, inferior parietal, fusiform cortices).

Results: Reconstruction across age groups. Averaged over the full lifespan cohort, MLP configurations achieved the lowest reconstruction error overall: jointly trained (0.347 ± 0.047 mm), fine-tuned (0.349 ± 0.048 mm), and directly applied (0.359 ± 0.050 mm) outperformed SUNet configurations. The MLP trained from scratch on Chinese data alone was by far the worst configuration overall (0.481 ± 0.070 mm), attributed to its large parameter count and lack of spatial inductive bias causing overfitting on only 640 subjects. SUNet configurations were tightly clustered (0.370–0.387 mm) regardless of adaptation strategy, and the jointly trained SUNet achieved the lowest SUNet error on Chinese data (0.370 ± 0.049 mm). Strikingly, the direct application of the UKB-trained models, without any exposure to Chinese data, generalizes surprisingly well across both architectures, even in pediatric and adolescent brackets (4–20 years) entirely outside UKB's age range, with only a mild increase in error at the youngest and oldest age brackets. Re-evaluation on UKB validation showed fine-tuning modestly increased UKB error (MLP: 0.256 → 0.285 mm; SUNet: 0.296 → 0.329 mm), while joint training improved UKB error relative to UKB-only training (MLP: 0.250 vs. 0.256 mm; SUNet: 0.281 vs. 0.296 mm). Scratch-trained models generalized back to UKB the worst, confirming that the SUNet-based SCSR model is more robust to a smaller training set than the MLP-based implementation.

Results: Atrophy detection. SUNet outperformed the MLP for all training configurations by an average AUC margin ranging from 0.066 (scratch) to 0.124 (joint training). Every configuration remained clearly above chance, indicating that all eight models learned a healthy reference that provides a viable basis for atrophy detection. Fine-tuning was the most effective strategy for SUNet, improving average AUC from 0.815 (direct) to 0.848 — the best result overall — and outperforming joint training (0.817). For the MLP, fine-tuning yielded only marginal improvement over direct application (0.732 vs. 0.720), and both scratch (0.699) and joint training (0.693) performed worse than direct application, suggesting adapting the substantially larger MLP requires more target-domain data than the 640 Chinese training subjects provide. Direct out-of-population transfer is described as a strong baseline across both architectures and the easiest adaptation strategy to deploy in practice.

Spatial patterns. Vertex-wise group mean Z-score maps from the fine-tuned SUNet model showed a clear, expected trend: AD patterns were particularly pronounced in temporoparietal and medial temporal regions, as well as in frontal areas, while MCI exhibited an intermediate pattern with more subtle, yet spatially consistent, cortical thinning in similar regions. Individual Z-score maps revealed substantial heterogeneity within diagnostic groups, underscoring the limitations of population-level averages for characterizing individual disease manifestations and the value of normative surface-based analysis.

Conclusion. The authors conclude that direct application of UKB-trained SCSR generalizes remarkably well to the Chinese cohort across nearly the entire lifespan, including age groups unseen during training. They offer a "practical, goal-dependent recommendation: fine-tuning is preferable when the priority is maximizing downstream diagnostic discrimination, while joint training is the better choice when the priority is retaining the best possible reconstruction accuracy across populations. Together with the strong generalization observed without target-domain adaptation, the results provide practical guidance for transferring SCSR to new international cohorts."

Improvements:

1. Architecture-aware transfer strategy selection. The improved system automatically chooses between a Spherical UNet (SUNet) and a large MLP based on target-domain dataset size. Given the finding that SUNet is robust to small training sets while the MLP overfits severely on only 640 subjects, the system selects SUNet when target-domain data is scarce and reserves the larger MLP for scenarios with abundant target data.

2. Goal-conditioned adaptation policy. Instead of a one-size-fits-all fine-tuning approach, the system distinguishes between two objectives:

• When maximizing downstream diagnostic discrimination (e.g., AD/MCI detection), it applies fine-tuning on target-domain healthy data, which improved SUNet AUC from 0.815 to 0.848.

• When maximizing reconstruction accuracy across populations, it uses joint training on pooled source + target healthy data, which achieved the lowest reconstruction error (MLP: 0.347 mm; SUNet: 0.370 mm).

3. Out-of-population zero-shot fallback. The system includes a direct-application mode for newly encountered cohorts with no local training data. It leverages the finding that UKB-trained models generalize surprisingly well even to age ranges completely unseen during training (e.g., 4–20 years), with only mild error increases at the extremes.

4. Age-stratified confidence estimation. The system estimates reconstruction reliability across the lifespan, flagging higher uncertainty at the youngest and oldest age brackets, allowing clinicians to interpret Z-scores with age-aware confidence.

5. Spatially-grounded diagnostic visualization. The improved system generates vertex-wise Z-score maps that highlight disease-specific patterns—temporoparietal, medial temporal, and frontal thinning for AD; more subtle, spatially consistent patterns for MCI—and surfaces individual heterogeneity, enabling personalized rather than purely population-average diagnosis.

6. Efficient multi-population normative modeling. The system supports a training regime where a single SUNet-based SCSR model is jointly trained on heterogeneous international cohorts, improving reconstruction on the original population (UKB error: 0.281 vs. 0.296) while also achieving near-best performance on the new population, reducing the need for separate per-cohort models.

What the improved AI system can do:

The improved system can deploy a normative cortical-thickness model to new international cohorts with minimal or no local training data, automatically selecting the optimal adaptation strategy based on the user’s goal—fine-tuning for maximal atrophy-detection accuracy or joint training for optimal reconstruction fidelity. It can accurately predict personalized cortical-thickness references across the human lifespan, including pediatric and adolescent subjects never seen during training, while quantifying age-dependent uncertainty. It provides vertex-wise Z-score maps that localize AD and MCI atrophy patterns with high discriminative power (SUNet fine-tuned AUC up to 0.848) and reveals individual-level disease heterogeneity. The system also avoids the common pitfall of overfitting in large-capacity models by preferring spatially-inductive spherical architectures when domain adaptation data is limited, and it retains robust performance across both source and target populations when jointly trained on pooled healthy data.

Stochastic cortical self-reconstruction (SCSR) enables personalized mapping of gray matter atrophy, a hallmark of neurodegenerative disorders such as Alzheimer's disease (AD), onto high-resolution cortical surfaces. Unlike conventional normative modeling approaches, which typically operate at a coarse regional level and remain inherently constrained by the covariates included during training, SCSR estimates an individualized healthy reference directly from the observed cortical thickness at the vertex level. This allows the detection of subtle, subject-specific deviations from healthy cortical shape. In this work, we investigate the generalization and transferability of SCSR, originally trained on UK Biobank (UKB) data, to an independent Chinese population dataset. Specifically, we evaluate the ability of SCSR-derived Z-scores to discriminate between healthy scans, individuals with mild cognitive impairment (MCI), and patients with AD, while also assessing model robustness across the lifespan. We compare four training strategies: direct application of the UKB-trained model, fine-tuning on Chinese data, training from scratch, and joint training on UKB and Chinese cohorts. As reconstruction backbones, we consider both a multilayer perceptron (MLP) and a Spherical UNet (SUNet). Our results demonstrate that SCSR provides robust detection of cortical atrophy in the Chinese population across all evaluated models. The highest discriminative performance was achieved by the fine-tuned SUNet model (average pairwise AUC = 0.848), followed closely by the UKB-trained SUNet. Moreover, reconstruction errors remained low across the lifespan, even when the training population exhibited a substantially narrower age distribution, indicating strong cross-population transferability.