Disentangling and Fusing Neurostructural and Vascular Ageing for Retinal Age Prediction

• The retinal age gap is associated with all-cause mortality, cardiovascular disease, and incident dementia, suggesting retinal age may reflect systemic ageing trajectories.Source: §I · p. 1
• Fundus-based retinal-age predictors usually produce one estimate and do not explicitly distinguish neurostructural change from vascular alteration or their combination.Source: §I · p. 1
• OCT and OCTA measurements can differ in reliability, motivating separate pathway estimates and fusion conditioned on reliability.Source: §I · p. 2
• SAP-DPF treats retinal ageing as neurostructural and vascular pathways plus a reliability-aware fused estimate, using OCT and OCTA.Source: §I · pp. 2–3, Figure 1 · p. 2
• 4.02 years MAE, a 12.52% improvement over the strongest single-modality baseline, was achieved by SAP-DPF on a six-modality OCT–OCTA dataset.Source: §I · p. 3

A. Age Prediction from Retinal Imaging

• Fundus and OCT studies established age-informative signals, but each estimates retinal age from a single modality without exploiting structural–vascular complementarity.Source: A. Age Prediction from Retinal Imaging · p. 3

B. Fusion Algorithms for Multimodal Prediction

• Le Goallec et al. used a fixed stacked combiner for fundus and OCT, while Pourjavan et al. used hand-crafted structural and vascular metrics.Source: B. Fusion Algorithms for Multimodal… · p. 3
• Fixed modality weights cannot track subject- and age-related reliability variation, such as motion artifacts degrading one scan or ageing cues shifting across the lifespan.Source: B. Fusion Algorithms for Multimodal… · p. 3
• Existing fusion strategies lack a weight that is simultaneously uncertainty-aware, age-adaptive, subject-specific, and attributable to a structural or vascular pathway.Source: B. Fusion Algorithms for Multimodal… · p. 3

III. METHODS

• The framework preserves separate structural and vascular age estimates before decision-level adaptive integration, allowing pathway-specific contributions to be examined.Source: §III · p. 3

A. Modality-Specific Age Regressors

• Six modalities are grouped into structural SDOCT-MC and CP pathways and vascular IVP, CC, SVC, and DVP pathways.Source: A. Modality-Specific Age Regressors · pp. 3–4, Figure 2 · p. 4
• Each modality uses a deep ensemble of five ResNet-50 regressors to produce age predictions with predictive variances.Source: Figure 2 · p. 4
• Independent modality regressors estimate age and aleatoric uncertainty; ensemble predictions additionally capture epistemic uncertainty due to limited training data.Source: A. Modality-Specific Age Regressors · p. 4
• Regressors use parallel heads for the predicted mean and log-variance and are trained with Gaussian negative log-likelihood against chronological age.Source: A. Modality-Specific Age Regressors · p. 4

B. Intra-group Fusion: Reliability-aware Fusion within OCT and OCTA

• The intra-group strategy uses decision-level predictions and uncertainty rather than fusing intermediate features across modalities.This avoids propagating vascular noise into a shared representation before modality-specific reliability can be assessed.Source: B. Intra-group Fusion:… · pp. 4–5
• The six modalities are partitioned into structural SDOCT-MC and CP, and vascular SVC, DVP, IVP, and CC groups before fusion.Predictions are aggregated within each group before the two pathways are merged.Source: B. Intra-group Fusion:… · p. 4
• Each modality's reliability prior is the reciprocal of a kernel-smoothed local prediction error, with bandwidth selected by leave-one-out cross-validation.The smoothing is intended to reflect reliability that varies smoothly with age; a small constant stabilizes the estimate.Source: B. Intra-group Fusion:… · pp. 4–5
• Within each group, modality predictions are fused using weights that combine age-dependent reliability priors with subject-specific predictive uncertainty.The priors capture population-level reliability variation with age, while inverse predictive variance down-weights modalities that are unreliable for an individual.Source: B. Intra-group Fusion:… · p. 5

C. Inter-group Fusion with Gated Mechanism

• The final prediction is a convex combination of structural and vascular pathway estimates, with their group-level uncertainties informing the gate.The gate uses a sigmoid of a low-order polynomial in age and a log-variance ratio, producing bounded, continuous weights.Source: C. Inter-group Fusion with Gated… · p. 5
• A second-order polynomial models slowly varying modality relevance, while the sigmoid keeps the pathway weighting continuous and bounded.The log-variance ratio modulates weights according to relative predictive uncertainty.Source: C. Inter-group Fusion with Gated… · p. 5

A. Dataset

• 504 subjects were split by subject into 352 training, 76 validation, and 76 test subjects, with 1,008 images per modality after quality control.The subset spans ages 30.2–83.2 years; mean age was 55.4 years (SD 14.7).Source: A. Dataset · p. 5
• Table II covers single-modality and intra-group fusion results on 152 test images from 76 subjects, with ΔMAE measured against each group's best single modality.The caption identifies circumpapillary as the structural comparator and DVP as the vascular comparator, but supplies no result values.Source: Table II · p. 5
• The structural fusion weights shift from SDOCT-MC dominance in younger subjects to CP dominance beyond approximately age 50.The caption describes these weights as derived from smooth reliability priors.Source: Figure 3 · p. 5
• Table III compares fusion strategies and ablations on the test set, including variants without the gate, dual-path grouping, or smooth prior.The caption defines subject MAE as predictions averaged across both eyes, but provides no table values.Source: Table III · p. 6
• Fig. 4 plots vascular-minus-structural prediction discrepancy against chronological age, shifting from positive in younger to negative in older individuals.The caption interprets this as age-dependent differences between the two pathway estimates.Source: Figure 4 · p. 6

B. Implementation Details

• Each modality uses an ImageNet-pretrained ResNet-50 deep ensemble of K = 5 members, trained with Adam and early stopping.Training uses learning rate 10−3, batch size 32, and patience 25.Source: B. Implementation Details · p. 6
• Fusion hyperparameters are selected by leave-one-out cross-validation, while remaining parameters are optimized on the validation set with regularization.The optimization minimizes mean absolute error; modality uncertainties are rescaled before weighting, followed by validation-set bias correction and variance calibration.Source: B. Implementation Details · p. 6
• The mean structural gating weight is π = 0.938, with a modest dip at ages 50–70 indicating a relative increase in vascular contribution.The age trend is shown with a LOWESS smooth and ±1 SD; the dashed line marks the global mean.Source: Figure 5 · p. 6
• Structural gating-weight distributions widen in the 70–80 and 80+ decades, suggesting greater variability in structural–vascular balance at older ages.The widening violin distributions are shown by age decade.Source: Figure 6 · p. 6

C. Prediction Performance

• SAP-DPF achieves the lowest overall MAE among evaluated fusion strategies (MAE=4.02 yr, R2=0.887) and significantly outperforms Simple Average, IVW-MAE, and IVW-Uncertainty.Stacking methods perform comparably, with ∆MAE≤0.21.Source: C. Prediction Performance · pp. 6–7
• SAP-DPF achieves the lowest observed NLL among fusion methods, significantly improving over four simpler methods, while showing small differences from strong stacking baselines.NLL differences are significant versus Stacking Ridge (p = 0.040) but not ElasticNet (p = 0.088); subject-level MAE is 3.79 versus 3.96 for the strongest stacking baseline.Source: C. Prediction Performance · p. 7
• In subject-level five-fold cross-validation, SAP-DPF wins 5/5 folds against Simple Average and IVW-MAE, 4/5 against IVW-Uncertainty, and 3/5 against ElasticNet.The authors report the pattern as consistent with the single-split comparison rather than an artifact of one partition.Source: C. Prediction Performance · p. 7
• For eyes aged ≥70 with above-median OCTA uncertainty, SAP-DPF reduces MAE by 5% versus ElasticNet (4.576 vs. 4.820), though larger-cohort confirmation is warranted.This subgroup contains n=16 eyes.Source: C. Prediction Performance · p. 7
• Gaussian 95% prediction intervals cover 94.1% of test eyes (143/152), and OCTA modality uncertainty correlates positively with prediction error.Across the four OCTA modalities, Pearson r reaches 0.34 and Spearman ρ reaches 0.44, consistent with the rationale for inverse-variance weighting.Source: C. Prediction Performance · p. 7

D. Clinical Interpretation

• The learned structural reliability prior shifts from favouring macular OCT in younger subjects to circumpapillary OCT in older subjects, with a crossover near age 50.This indicates age-dependent predictive reliability differences between the two structural modalities in this cohort.Source: D. Clinical Interpretation · p. 7
• Structural predictions remain the primary driver across the cohort (π = 0.938), while a midlife reduction and increased variability among older subjects indicate variation in vascular contribution.The reduction occurs at ages 50–70, and the older-subject variability is reflected in Fig. 6.Source: D. Clinical Interpretation · p. 7
• Structural–vascular prediction discrepancy shifts from positive in younger subjects, where vascular predicts older, to negative in older subjects.The authors describe this as different age-dependent prediction patterns across structural and vascular imaging.Source: D. Clinical Interpretation · p. 7
• Structural OCT provides the more accurate age estimate, whereas vascular age variability may reflect imaging limitations and systemic cardiovascular risk factors.The authors propose assessing whether vascular age gaps provide complementary cardiovascular-risk information.Source: D. Clinical Interpretation · p. 7

V. CONCLUSION

• SAP-DPF separately estimates structural age from OCT and vascular age from OCTA before integrating the predictions, with structural OCT more accurate and vascular predictions modestly improving the fused MAE to 4.02 years.Source: §V · p. 7
• Smooth age-dependent reliability priors and sample-level uncertainty support intra-group fusion, followed by an uncertainty-aware gate for subject-specific integration.The design preserves pathway-specific age estimates and interpretable fusion weights for examining structural–vascular prediction differences.Source: §V · pp. 7–8
• The biological significance of the pathway-specific outputs requires further validation.Source: §V · p. 8
• The study uses a single cohort of moderate size, so generalization across populations, imaging devices, and acquisition protocols remains unestablished.Source: §V · p. 8