Disentangling and Fusing Neurostructural and Vascular Ageing for Retinal Age Prediction

summary

The gist

Estimating biological age is an important task in ageing research, as it quantifies individual ageing trajectories beyond chronological age<ref:2609.38264#pg2> Retinal age estimation has become a…

In short

This work develops a framework to estimate biological age from retinal images by separately modeling neurostructural ageing using OCT and vascular ageing using OCTA. The proposed SAP-DPF framework uses deep ensembles and a reliability-aware fusion method to combine these estimates, achieving high accuracy while revealing how modality reliability changes with age.

Key concepts

Retinal Age Gap
This measures the difference between the predicted retinal age from an image and a person's actual chronological age. This gap is clinically significant as it has been linked to increased risks for all-cause mortality, cardiovascular disease, and dementia.
Deep Ensemble Regressors
The model uses multiple independent ResNet-50 neural networks (a deep ensemble) for each imaging modality. These regressors are trained to predict age and also estimate uncertainty. This helps the system understand both the noise inherent in the images (aleatoric uncertainty) and the limitations of the training data itself (epistemic uncertainty).
Reliability-Aware Fusion
The framework combines structural and vascular age predictions using a weighted scheme. The weights are determined by a gate mechanism that considers both the proxy age and how confident each individual pathway is in its prediction, allowing for an adaptive fusion strategy.
Structural-Vascular Discordance
This refers to the pattern where the relationship between structural and vascular age estimates changes over time. The study found this discordance shifts from vascular predictions being older in younger subjects to structural predictions being more accurate in older subjects.

Terminology used across episodes

This episode discusses

The paper

Disentangling and Fusing Neurostructural and Vascular Ageing for Retinal Age Prediction · Read on arXiv

Junwen Zheng, Li Rong Wang, Anthony Zihan Lin, Xinran Xu, Wei Kiong Ngo, Kelvin Zhenghao Li, Tock Han Lim

Nanyang Technological University · Centre for Frontier AI Research, A*STAR

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Disentangling and Fusing Neurostructural and Vascular Ageing for Retinal Age Prediction".

Dev: Here's the gist of the paper. Estimating biological age is an important task in ageing research, as it quantifies individual ageing trajectories beyond chronological age<ref:2609.38264#pg2> Retinal age estimation has become a well-established direction in this area because retinal imaging provides a non-invasive window into neural and microvascular ageing<ref:2609.38264#pg2> This work fills this gap by studying neurostructural ageing with Optical Coherence Tomography (OCT) images and vascular ageing with Optical Coherence Tomography Angiography (OCTA) images<ref:2609.38264#pg2> The proposed SAP-DPF framework achieves a mean absolute error of 4.02 years, representing a 12.52% improvement over the strongest single-modality baseline<ref:2609.38264#pg2> This research provides a framework for investigating structural and vascular contributions to heterogeneous retinal ageing<ref:2609.38264#pg2>.

Rosa: Pruning redundant reasoning tokens Retinal ageing is a clinically meaningful biological ageing phenotype because the retina captures both neural and microvascular changes through non-invasive imaging<ref:2609.38264#pg2> The retinal age gap, defined as the difference between image-predicted retinal age and chronological age, has been associated with all-cause mortality

1: , cardiovascular disease

2: , and incident dementia

3: <ref:2609.38264#pg2> Existing studies have predominantly focused on fundus photographs and usually model retinal ageing as a single generic process<ref:2609.38264#pg2> Little research has explicitly distinguished neurostructural and vascular ageing in retinal age prediction<ref:2609.38264#pg2>.

Dev: How the training signal is built The framework treats retinal ageing as three related outputs rather than a single fused prediction<ref:2609.38264#pg2> The design reflects the medical premise that neurostructural and vascular ageing may encode distinct biological information and may differ in reliability across subjects and ages<ref:2609.38264#pg2>.

Rosa: The model employs a deep ensemble (K=5) of ResNet-50 regressors for each modality to produce age predictions with predictive variances<ref:2609.38264#pg2> Each regressor explicitly estimates heteroscedastic aleatoric uncertainty, capturing modality- and sample-specific noise characteristics<ref:2609.38264#pg2>.

Title and authors: Rosa: We're talking about the title of this paper again, "Disentangling and Fusing Neurostructural and Vascular Ageing for Retinal Age Prediction." The authors are Zheng, Wang, Lin, Xu, Tan, Li, Lim, and Fan from Nanyang Technological University Singapore.

Dev: That’s a whole team behind it. It sounds pretty technical but the core idea is separating the structural changes in the retina from the vascular changes.

Rosa: Right. What this paper does is take those two separate estimates—the structural one and the vascular one—and then fuse them in a way that respects how reliable each estimate is for a specific person and age.

Dev: It’s not just averaging them; it uses something called a reliability-aware fusion mechanism. They’ve designed it so that if one pathway is less trustworthy at a certain age, the final prediction leans more on the other one.

Rosa: That’s the key part, right? It acknowledges that neurostructural and vascular ageing might give different kinds of information, and they have to account for that difference.

Dev: And they're using deep ensembles of ResNet-fifty regressors for each modality, which helps them estimate not just the age but also how much uncertainty there is in the prediction itself.

Rosa: So, it’s not just a single number; it’s a set of predictions with associated uncertainties for both structures and vessels.

Dev: That uncertainty part is important because it tells you when the model might be less sure about what it's seeing on that specific scan.

The paper's summary: Rosa: So, to summarize what they did, this paper formulates multimodal OCT and OCTA age estimation as a decomposition problem between structural and vascular ageing.

Dev: They then instantiate this using a framework called SAP-DPF, which is this dual-path approach that fuses the two pathways based on predictive reliability.

Rosa: The main contribution they highlight is moving away from that single overall retinal age estimate you see in older fundus photos to getting these two separate estimates.

Dev: They also introduce smooth reliability priors that change based on the age, which lets them model how much more reliable one estimate is compared to the other as a person gets older.

Rosa: That’s what they mean when they say the design reflects the medical premise that these two types of ageing might encode different biological information.

Dev: And because vascular ageing estimates can be noisier than structural ones, their final fusion strategy is designed to be age-adaptive and subject-specific.

The paper's improvements: Rosa: They suggest a few specific ways this SAP-DPF framework improves upon previous methods. First, they use a reliability-weighted scheme for intra-group fusion where the weight depends on the prediction variance of each pathway and the age itself.

Dev: And for inter-group fusion, they use a gated mechanism that balances the two pathways using both the proxy age and the log variance ratio as inputs.

Rosa: That gating mechanism is smart because it uses individual-level prediction confidence to decide how much to trust which pathway at any given time.

Dev: They also show how this reliability prior shifts across subjects, with their learned reliability prior moving from favoring macular OCT in younger subjects toward circumpapillary OCT in older subjects, with a crossover near age fifty.

Rosa: That shift is interesting because it shows how the model adapts its trust based on the subject’s age and location in the eye.

Dev: And when they look at the results, they find that while the structural pathway remains the primary driver across the whole cohort, there’s a systematic discrepancy between them.

Conclusion: Rosa: So, wrapping this up for "Disentangling and Fusing Neurostructural and Vascular Ageing for Retinal Age Prediction," it shows that we can do more than just giving one number. We can get a structural age estimate and a vascular age estimate separately.

Dev: The paper’s main implication is that by using this SAP-DPF framework, you get an improved accuracy of four point zero two years in mean absolute error compared to the best single-modality baseline.

Rosa: That’s a twelve point five two percent improvement over what was previously available. It gives us a concrete method for investigating how structural and vascular changes contribute separately to retinal ageing.

Dev: The final result is that the structural OCT provided the more accurate age estimate, while the variability in vascular age seems to reflect both imaging limitations and systemic cardiovascular risk factors.

Rosa: It’s a solid framework for anyone looking into the different biological drivers of retinal ageing. We’ll keep an eye on how this decomposition approach evolves.

Dev: Agreed. The paper provides a detailed blueprint for how to build these reliability-aware, subject-specific fusion models in the future.

More episodes

← Home