💻 computer science

Modelling Geographic Atrophy Progression using Implicit Neural Representations

This paper proposes a low-data Implicit Neural Representation (INR) framework to model individual Geographic Atrophy progression in Age-related Macular Degeneration, achieving superior lesion area prediction and segmentation accuracy while maintaining high-quality Fundus Autofluorescence image generation across past and future time points.

Original authors: Simone Sarrocco, Paul Friedrich, Florentin Bieder, Christina Bornberg, Philippe Valmaggia, Peter Maloca, Philippe Cattin

Published 2026-08-12

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Simone Sarrocco, Paul Friedrich, Florentin Bieder, Christina Bornberg, Philippe Valmaggia, Peter Maloca, Philippe Cattin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how a drop of ink will spread across a piece of wet paper. You can see the ink at the start, and you can see it a little later, but the exact shape it takes in between, or where it will be tomorrow, is a mystery. This is the kind of puzzle scientists face when studying a specific type of eye disease called Age-related Macular Degeneration (AMD). In its late, "dry" stage, this disease creates a patchy, growing area of damage in the back of the eye called Geographic Atrophy (GA). Doctors currently use special cameras to take pictures of the eye's back (called Fundus Autofluorescence or FAF) to see how these dark patches grow over time. However, because every person's eye is different and the disease moves at its own unique speed, it's hard to predict exactly how a specific patient's vision will change. The goal is to build a digital crystal ball that can look at a few past photos and guess what the eye will look like in the future, helping doctors explain to patients what might happen next.

This paper introduces a clever new way to build that crystal ball using something called "Implicit Neural Representations" (INRs). Think of an INR not as a standard photo album, but as a magical, continuous recipe book. Instead of storing a picture as a grid of pixels (like a digital photo), this recipe book stores a set of instructions that can describe the image at any point in time or space. If you ask the recipe, "What does the eye look like at week 5?" it calculates the answer on the fly, even if no one ever took a photo at week 5. The researchers used this method to create a model that learns the unique "story" of a patient's eye disease. They found that by giving the model a special "identity card" (a latent vector) for each specific eye, it could learn the unique shape and growth pattern of that person's damage.

The team tested their model on a group of patients who had been scanned up to four times over several months. They asked the model to do two things: first, to recreate the blurry, glowing pictures of the eye (the FAF images), and second, to draw a precise outline of the damaged area (the GA segmentation). The results were quite promising. In tests where the model had to guess the future or fill in missing time points, it was surprisingly good at predicting the size and shape of the damaged area. It achieved the highest accuracy in measuring the lesion size and the best match for the shape of the damage compared to other computer programs. While it wasn't perfect at recreating the fine, high-definition details of the glowing eye pictures (sometimes just copying the last known picture was actually better for image clarity), it excelled at understanding how the disease was changing. The authors suggest this approach could help doctors visualize a patient's specific disease journey, potentially making it easier to discuss future outcomes and treatment benefits. However, they note that more work is needed to make the image details sharper and to test this on a larger group of people with different stages of the disease.

Technical Summary: Modelling Geographic Atrophy Progression using Implicit Neural Representations

Problem Statement
Age-related Macular Degeneration (AMD) is a leading cause of blindness, with its late dry phase characterized by Geographic Atrophy (GA), an irreversible loss of retinal tissue. While Fundus Autofluorescence (FAF) imaging is the standard for monitoring GA lesion growth, the highly individualized nature of disease progression makes it difficult to model and predict evolution accurately, particularly in low-data settings where longitudinal patient records are sparse. Existing deep learning approaches often struggle to capture continuous, patient-specific trajectories or require extensive data to generalize.

Methodology
The authors propose a framework utilizing Implicit Neural Representations (INRs) to model GA progression as a continuous, subject-specific function. The core architecture is an auto-decoder consisting of a Multi-Layer Perceptron (MLP) with SIREN activation functions.

Architecture and Conditioning: The model is split into two heads: a reconstruction head ( $f^{faf}_\theta$ ) for generating FAF image pixel intensities and a segmentation head ( $f^{seg}_\theta$ ) for predicting GA lesion masks. Both heads share the same MLP up to the penultimate layer.
Latent Representation: To capture individual eye characteristics, a unique 3D spatial latent vector ( $z_i$ ) is assigned to each eye. This vector is shared across all time points for that specific eye, forcing the model to learn eye-specific anatomical features rather than visit-specific artifacts.
Modulation and Temporal Dynamics: The network is conditioned via modulation layers (FiLM). The latent vector is concatenated with temporal variables (weeks elapsed since baseline) and clinical variables (patient age at the visit). These combined inputs are linearly mapped to scale and shift parameters for each layer. Notably, the shift parameter is used to represent temporal dynamics as low-level signals.
Training and Test-Time Adaptation:
- Training: The model optimizes both network parameters ( $\theta$ ) and the set of eye-specific latent vectors ( $\{z_i\}$ ) by minimizing a joint loss function comprising Mean Squared Error (MSE) for FAF reconstruction and a combination of Binary Cross Entropy and DICE loss for segmentation.
- Test-Time Adaptation: For previously unseen patients (new eyes), a new latent vector is randomly initialized and optimized using the available visits for that specific patient while keeping the network weights frozen. This allows the model to adapt to individual disease trajectories without retraining the entire network.

Key Contributions

First Application of INRs for GA: This work presents the first application of Implicit Neural Representations to model individual lesion trajectories in GA secondary to AMD.
Joint Image and Segmentation Modeling: The framework simultaneously generates FAF images and GA segmentation masks at arbitrary time points (past, present, and future), enabling interpolation and extrapolation of disease states.
Low-Data Adaptability: By employing test-time adaptation of eye-specific latent vectors, the method generates accurate predictions for new patients using limited longitudinal data (as few as one or two visits).
Continuous Trajectory Modeling: The approach treats disease progression as a continuous process, allowing for the visualization of intermediate states and future evolution that discrete models might miss.

Results
The method was evaluated on the OMEGA dataset, comprising 37 eyes from 30 patients with up to four visits. Comparisons were made against classical interpolation methods, ImageFlowNet, T-UNet, and a time-aware diffusion model (T-I2SBUNet) across three scenarios: single image pair prediction, full history extrapolation, and missing visit reconstruction.

Segmentation Performance: The proposed method achieved the highest DICE scores (0.85 in Scenario 1, 0.91 in Scenario 2) and the lowest Mean Absolute Error (MAE) for lesion area (0.31 mm² and 0.20 mm², respectively) across all scenarios. It significantly outperformed other deep learning models in capturing lesion shape and localization.
Image Reconstruction: The model achieved competitive PSNR and SSIM scores for FAF reconstruction. However, the authors note that in some cases, simply copying the last available ground truth image yielded the best reconstruction metrics, likely due to the short follow-up intervals (approx. 12 weeks) in the dataset.
Qualitative Analysis: Visual inspection confirmed that the model generates realistic lesion growth trajectories that align with ground truth measurements, whereas some competing methods produced low MAE scores but visually mismatched atrophic regions.

Significance and Claims
The paper claims to demonstrate the utility of spatial latent vectors for learning individual anatomical structures and injecting time information via modulation layers. The authors position the work as a tool to help clinicians visualize expected disease evolution and communicate potential treatment benefits to patients.

The authors remain modest regarding the current limitations, acknowledging that reconstructing high-frequency details in FAF images during extrapolation remains a challenge. They emphasize that while the model performs well in segmentation and area estimation, image quality in extrapolation scenarios requires further study. The paper concludes that future work should focus on larger cohorts, different AMD stages, and the integration of additional modalities (e.g., OCT) to improve clinical interpretability and risk prediction.

Technical Summary: Modelling Geographic Atrophy Progression using Implicit Neural Representations

More like this