PaperlayerView exact revision on arXiv

Source-linked AI summary

Systematic Morphometric Brain Autoencoding for Representation Learning in Neurodegenerative Diseases

Nuria Balbás-Esteban, Albert Belenguer-Llorens, Elaheh Moradi, Jussi Tohka, Vanessa Gómez-Verdejo, Carlos Sevilla-Salcedo

arXiv:2610.37074v1cs.CVcs.LG

TL;DR

Structural MRI offers a way to study distributed brain changes in neurodegeneration, but annotations are scarce and existing self-supervised objectives may not capture those morphometric patterns. SYMBA uses anatomically constrained masking, structural-similarity reconstruction, and a progressive curriculum; on unseen cohorts it outperforms existing SSL approaches, including in a six-subject labeling regime. The evaluation relies on frozen linear probing, and the study notes that this does not establish the maximum performance achievable through adaptation.

  • Problem

    Generic natural-image masking and voxel-wise reconstruction may not capture the distributed morphometric patterns of neurodegeneration, while diagnostic annotations are scarce and supervised models can be sensitive to distribution shifts.

    Source: §1 · p. 2

  • Method

    SYMBA combines anatomically constrained masking, MS-SSIM reconstruction, and a progressive curriculum that shifts from local detail to global brain morphology.

    Source: §1 · p. 3

  • Results

    SYMBA showed superior performance and out-of-domain generalizability across clinical tasks on two unseen datasets, reaching approximately 0.90 AUROC for Alzheimer’s disease classification with six training samples.

    Source: §5 · p. 22

  • Takeaways & Limitations

    The findings support incorporating neuroanatomical priors into self-supervised objectives to learn transferable structural MRI representations and reduce dependence on large annotated cohorts.

    Source: §5 · p. 22

    Frozen linear probing assesses representation quality but does not establish maximum predictive performance achievable through adaptation.

    Source: §4.4 · p. 21

Abstract

from arXiv · show

Early identification of neurodegenerative disorders using structural T1-weighted MRI is essential for timely clinical intervention, yet supervised deep learning models often struggle to generalize due to limited annotations and scanner-induced distribution shift. Self-supervised learning offers a promising alternative; however, existing approaches are largely adapted from natural image paradigms, relying on contrastive objectives or generic masked reconstruction with voxel-wise losses, which may not fully capture the distributed morphometric patterns associated with neurodegeneration. We introduce SYMBA (SYstematic Morphometric Brain Autoencoding), a self-supervised pretraining framework specifically designed for structural brain MRI. SYMBA integrates anatomically constrained masking, a multi-scale structural similarity (MS-SSIM) reconstruction objective, and a progressive curriculum that transitions from local anatomical detail to global brain morphology. These components explicitly incorporate neuroanatomical structure into the self-supervised objective rather than relying on generic image priors. Pretrained on 6,953 T1-weighted MRI scans from nine heterogeneous cohorts, SYMBA is evaluated using frozen linear probing on two independent unseen datasets, enabling a strict assessment of out-of-distribution generalization. Across multiple clinical tasks, SYMBA consistently outperforms recent neuroimaging foundation models, achieving 0.930 AUROC for Alzheimer's disease classification, 0.772 for mild cognitive impairment conversion prediction, and strong performance in amyloid status prediction. In low-label regimes, SYMBA demonstrates high sample efficiency, reaching approximately 0.90 AUROC for AD/CN classification using only six labeled subjects.

1. Introduction

Structural MRI captures subtle, progressive changes distributed across brain networks, but neurodegenerative disease analysis is constrained by scarce annotations and distribution shifts. SYMBA addresses gaps in self-supervised objectives with anatomy-aware reconstruction and evaluates transfer from healthy subjects to unseen neurodegenerative cohorts.

1. Introduction

  • Neurodegenerative changes are subtle and distributed across interconnected brain networks, while supervised models face scarce annotations and scanner- and population-related distribution shifts.Structural MRI offers volumetric depictions of brain anatomy for learning transferable representations.Source: §1 · p. 2
  • Structural segmentation and registration prioritize anatomical localization, whereas neurodegeneration representation learning must capture distributed morphometric alterations and long-range relationships.Relevant changes can extend across interconnected regions as disease evolves.Source: §1 · pp. 2–3
  • Existing neurodegeneration SSL studies emphasize contrastive learning, while generic masking and voxel-wise losses do not explicitly model brain anatomy or its spatial correlations.The relative advantages of contrastive and generative objectives remain underexplored, and end-to-end fine-tuning can confound evaluation of learned representations.Source: §1 · p. 3
  • SYMBA combines anatomically constrained masking, MS-SSIM reconstruction, and a progressive curriculum from local detail to global morphology.The framework is designed to learn morphometric representations for downstream neurodegenerative disease analysis.Source: §1 · p. 3
  • SYMBA is pretrained predominantly on cognitively healthy subjects and evaluated against recent SSL methods on two unseen neurodegenerative cohorts using frozen representations and low-label regimes.The evaluation is framed as a controlled assessment of representation quality, transferability, and robustness under distribution shift.Source: §1 · pp. 3–4

2. Methodology

SYMBA learns structural-MRI representations by masking brain anatomy and reconstructing missing regions, then uses the retained encoder features for frozen linear probing.

2.1. Framework Overview

  • Masked structural MRI volumes are encoded into latent representations and decoded to reconstruct the original images.The framework is trained on unlabeled scans.Source: §2.1 · pp. 4–5

2.2. Anatomically Constrained Masking Operator

  • SYMBA restricts geometric masks to brain tissue, mirrors selected regions across hemispheres, and progressively increases masking scale and region count.The curriculum shifts the task from local inpainting toward reconstruction of large-scale neuroanatomical structures.Source: §2.2 · pp. 5–6

2.3. Reconstruction Objective

  • Reconstruction loss is computed only over masked voxels and uses MS-SSIM to compare luminance, contrast, and structure across spatial resolutions.This avoids supervision from visible voxels and homogeneous background regions.Source: §2.3 · p. 7

2.4. Downstream Tasks

  • After pretraining, SYMBA discards the decoder and evaluates fixed encoder embeddings with logistic-regression linear probes.The frozen protocol isolates the discriminative information in learned representations from downstream model capacity.Source: §2.4 · p. 7

3.1. Materials

SYMBA was pretrained on a heterogeneous collection of structural MRI scans, while ADNI and AIBL were reserved for downstream evaluation. The evaluation covered diagnostic, prognostic, amyloid, and sex-classification tasks using baseline T1-weighted scans.

3.1. Materials

  • 6,953 T1-weighted scans from seven cohorts formed a heterogeneous pretraining corpus spanning scanners, acquisition protocols, and subject populations.The corpus was predominantly cognitively healthy, with OASIS as the exception.Source: §3.1 · p. 8
  • ADNI and AIBL were held out for downstream evaluation, with baseline T1-weighted scans used to assess out-of-distribution generalization.The two cohorts were not used during pretraining.Source: §3.1 · p. 8
  • Tasks covered AD versus cognitively normal classification, MCI progression, three-class diagnosis, amyloid status, and sex classification.Stable MCI required five years without progression; progressive MCI required dementia diagnoses at the final two consecutive evaluations.Source: §3.1 · p. 9
  • Only AD versus CN, three-class diagnosis, and sex classification included both ADNI and AIBL; MCI progression and amyloid tasks used ADNI only.Structural MRI inputs were stereotactically normalized gray-matter-density maps, then resized to 128×160×128 voxels and intensity-scaled to [0,1].Source: §3.1 · pp. 9–10

3.2. Experimental setup

SYMBA uses a 3D ResNet-50 encoder and a transposed-convolution decoder, pretrained with a progressively harder masking curriculum and masked-region structural-similarity reconstruction. Frozen representations are evaluated with nested cross-validation and logistic-regression probes.

3.2. Experimental setup

  • SYMBA uses a 3D ResNet-50 encoder producing a 2,048-dimensional embedding and a decoder made of five transposed-convolution layers followed by sigmoid activation.Source: §3.2 · p. 10
  • Across 600 pretraining epochs, the masking ratio rose from 30% to 75%, while characteristic scale shifted from 16 to 32 after epoch 300.Pretraining used healthy-control cohorts and OASIS, excluding ADNI and AIBL.Source: §3.2 · p. 10
  • AdamW optimization used weight decay 0.05, peak learning rate 1×10^-4, and batch size 4 via gradient accumulation; logistic-regression probes were tuned with nested 5-fold/3-fold cross-validation.The code was publicly available, but the datasets were not hosted in the repository because of licensing constraints.Source: §3.2 · pp. 10–11
  • Reconstruction used a four-scale masked-region 3D MS-SSIM loss, considering only masked brain regions with ρ ≥ 0.20 and excluding background voxels.Source: §3.2 · p. 10

3.3. Performance evaluation

Under frozen-representation evaluation, SYMBA achieved the highest BACC and AUROC across classification tasks on both ADNI and AIBL. Most improvements were statistically significant, with the sMCI-versus-pMCI advantage narrower and not significant.

3.3. Performance evaluation

  • The comparison evaluated representative self-supervised models as frozen feature extractors using the same linear-probing protocol; AIBL provided the strictest out-of-distribution test among the compared models.BrainMorph, AnatCL, and BrainIAC had used ADNI in pretraining, whereas BrainSegFounder and SYMBA had not.Source: §3.3 · p. 11
  • SYMBA achieved the best BACC and AUROC on every classification task across both datasets, with most improvements statistically significant.The margin over the strongest baselines narrowed and lost significance on the more challenging sMCI-versus-pMCI task.Source: §3.3 · p. 11, Table 3 · p. 11
  • On ADNI, AD-versus-CN performance rose from 0.82 to 0.87 BACC and 0.90 to 0.94 AUROC over the best baseline; SYMBA also achieved 0.71 BACC and 0.78 AUROC on sMCI versus pMCI.Improvements also extended to three-class diagnosis and amyloid prediction.Source: §3.3 · p. 11
  • On AIBL, SYMBA again achieved the highest BACC and AUROC across tasks, with significant differences over most baselines.Source: §3.3 · p. 11

3.4. Low-data regime evaluation

Low-label evaluation compared SYMBA with supervised-from-scratch and pretrained-model baselines as training data were reduced. SYMBA generally led in sparse-label settings, with different patterns on sMCI-versus-pMCI and across the independent AIBL cohort.

3.4. Low-data regime evaluation

  • Training sets were subsampled to 1%, 2%, 8%, 25%, 50%, 75%, and 100%, with a fixed test partition and ten repeats per fold.Baselines were the same architecture trained from scratch and four pretrained neuroimaging models evaluated with frozen encoders and linear probes.Source: §3.4 · p. 12
  • On ADNI, SYMBA outperformed the supervised baseline and four pretrained models on AD-versus-CN, sex, and amyloid classification, especially with few labels.With five to six labeled participants, it already outperformed all four pretrained models on AD versus CN; the supervised baseline caught up at full data, while the other models plateaued lower.Source: §3.4 · p. 12, Figure 3 · p. 13
  • For sMCI versus pMCI, SYMBA led in the extreme low-data regime and remained ahead of the four pretrained models, but the supervised baseline overtook it at moderate label amounts.Source: §3.4 · p. 12
  • On AIBL, SYMBA remained top-performing across label fractions, and the supervised baseline did not catch up at full data; pretrained models trailed both methods.BrainMorph had seen AIBL during pretraining but still trailed SYMBA.Source: Figure 3 · p. 13, §3.4 · p. 14

3.5. Ablation study

Controlled ablations isolate the effects of architecture, pretraining paradigm, masking constraints, and reconstruction loss. The anatomical shape prior produced the broadest and most consistent degradation when removed.

3.5. Ablation study

  • The ablation study varied one design axis at a time while holding the remaining training configuration fixed to assess effects on representation quality and transferability.The three axes were autoencoder architecture, pretraining paradigm, and reconstruction objective; masking constraints were also examined through leave-one-out variants.Source: §3.5 · p. 14, §3.5.2 · pp. 15–16

3.5.1. Autoencoder architecture

  • On ADNI, ResNet L led across all five tasks, while the larger UNet L was weakest on most tasks; on AIBL, UNet L led AD vs CN BACC and was competitive on 3-Class.The architecture comparison included residual models of different capacities and a UNet with skip connections.Source: §3.5.1 · pp. 14–15, Table 4 · p. 14

3.5.2. Masking strategy and SSL paradigm

  • The full shape-prior objective was best or joint-best on most tasks across both cohorts; MAE modestly improved over contrastive baselines but consistently trailed SYMBA.AIBL amyloid and 3-class discrimination were exceptions, where plain SimCLR remained competitive with, or marginally ahead of, the shape-prior objective.Source: §3.5.2 · p. 15, Table 5 · p. 15
  • Removing the anatomical shape prior caused the broadest, most consistent degradation, with significant drops including ADNI Sex AUROC from 0.94 to 0.89 and AIBL Sex AUROC from 0.92 to 0.88.Turning off background removal, hemispheric symmetry, and minimum-tissue-fraction constraints had smaller, selective effects; several AIBL results were on par with or nominally exceeded the full model.Source: §3.5.2 · p. 16, Table 6 · p. 16

3.5.3. Reconstruction objective

  • Across cohorts, MS-SSIM generally favored transfer, while no reconstruction loss led on every task and single-scale SSIM led selected AIBL AUC results.MSE led AUC for ADNI MCI progression and amyloid status; SSIM+L1 did not consistently improve over either individual loss, aside from AIBL Sex BACC.Source: §3.5.3 · pp. 16–17, Table 7 · p. 17

3.6. Reconstruction analysis

Under 75% masking, the final SYMBA model reconstructed representative held-out volumes with global anatomical structure and major boundary continuities. Figure 4 lays out the original, masked, reconstructed, and difference views.

3.6. Reconstruction analysis

  • Despite 75% masking, the model recovered overall anatomical structure and major boundary continuities in randomly selected held-out examples.The authors use the visualizations to inspect information encoded in the latent space, while quantitative evaluation focuses on representation transferability.Source: §3.6 · p. 17
  • The illustrated subjects were randomly selected from the held-out test set without prior visual inspection.Source: §3.6 · p. 17
  • Figure 4 presents axial and coronal views reconstructed with ResNet using 75% masking and patch size 16.Its columns show the original image, masked input, reconstruction, total absolute difference, and masked difference.Source: §3.6 · p. 17, Figure 4 · p. 18

4. Discussion

The discussion links SYMBA’s transferability to reconstruction objectives that preserve spatial relationships and global morphology, while emphasizing that usefulness varies by task and evaluation setting. It also identifies limits in the cohorts, modalities, and downstream adaptation assessed.

4.1. Learning Morphological Representations for Neurodegenerative Disease Analysis

  • Reconstructing missing tissue from anatomical context may retain spatial relationships, while progressive masking presents reconstruction problems at multiple scales.The discussion contrasts this objective with instance discrimination, anatomical correspondence, and segmentation.Source: §4.1 · p. 19
  • Improvements were less pronounced for MCI progression and amyloid prediction, where macroscopic morphology may represent clinical progression and molecular pathology less reliably.The discussion frames structural MRI as an indirect characterization of underlying pathology.Source: §4.1 · p. 19

4.2. Transferability and Label Efficiency

  • Frozen linear probes on independent cohorts support SYMBA’s transferability, but do not establish the maximum performance achievable after fine-tuning.The protocol tests linear accessibility of downstream information rather than each architecture’s full potential after adaptation.Source: §4.2 · pp. 19–20
  • SYMBA’s advantage was most pronounced with very few labels, when information for downstream discrimination was already accessible in the pretrained feature space.This may matter where labels require expensive biomarker measurements, specialist assessments, or prolonged follow-up.Source: §4.2 · p. 20
  • Pretraining benefits were not uniform: supervised models approached SYMBA on several ADNI tasks with more labels, and eventually surpassed frozen representations for MCI progression.The discussion suggests task-specific feature combinations may not be linearly accessible in the pretrained space; fine-tuning may be useful when more annotations are available.Source: §4.2 · p. 20

4.3. Anatomical Inductive Biases and Representation Quality

  • Skip connections may bypass the latent bottleneck, but U-Net performance on some endpoints also suggests fine-grained spatial detail can help.The discussion cautions that skip connections are not universally detrimental and that reconstruction fidelity need not align with downstream representation quality.Source: §4.3 · p. 20
  • Geometrically structured masking and multiscale structural similarity offer useful supervision, but anatomical constraints and reconstruction losses may favor different anatomical information.The discussion recommends balancing anatomical detail, spatial context, and latent compression rather than maximizing capacity or reconstruction fidelity alone.Source: §4.3 · pp. 20–21

4.4. Limitations and Future Directions

  • The evaluation does not isolate scanner, acquisition, or population effects; it also relies mainly on frozen linear probing and baseline T1-derived gray matter maps, without modeling within-subject change.Some competing models used evaluation-cohort data; the discussion calls for controlled acquisition-shift studies and notes that complementary modalities or longitudinal data could address current scope boundaries.Source: §4.4 · p. 21

5. Conclusions

The paper presents SYMBA as anatomically informed generative pretraining for structural brain MRI and reports superior performance across clinical tasks on two unseen external datasets. In a low-data setting, it reports approximately 0.90 AUROC for Alzheimer’s disease classification using six training samples.

5. Conclusions

  • SYMBA combines neuroanatomical priors in its masking strategy, reconstruction objective, and training curriculum to model brain structure across spatial scales.The framework is presented as self-supervised pretraining systematically optimized for structural brain MRI.Source: §5 · p. 22
  • SYMBA outperformed existing SSL approaches for structural MRI and showed out-of-domain generalizability across clinical tasks on two independent unseen datasets.The evaluation used frozen linear probing on ADNI and AIBL.Source: §5 · p. 22
  • Approximately 0.90 AUROC for Alzheimer’s disease classification was achieved with six training samples.The conclusion describes these six samples as 1% of the training data and reports that SYMBA outperformed fully supervised baselines trained from scratch.Source: §5 · p. 22
Loading 2610.37074v1…