AI for Age-Related Macular Degeneration Assessment

This review highlights the increasing prevalence of AMD due to an aging population and the growing burden on healthcare systems. Concurrent advances in imaging modalities, particularly Optical Coherence Tomography (OCT), provide vast amounts of patient data. AI is presented as a crucial tool to address these challenges by providing fast, objective, and reproducible assessment of AMD across all stages. The article details how AI algorithms, specifically Machine Learning (ML) and Deep Learning (DL), are being developed and applied for classification, segmentation, and prediction of AMD progression, particularly focusing on early/intermediate AMD, geographic atrophy (GA), and neovascular AMD (nAMD). While acknowledging the significant benefits of AI, the authors also discuss current limitations and future directions.

Key Themes and Important Ideas/Facts:

Increasing Burden of AMD and Data Volume:

The global population is aging, leading to a rise in age-dependent diseases like AMD.
Estimates project 288 million people affected by AMD in 2040.
The increasing number of patients, visits, and treatments, coupled with advancements in multimodal imaging generating large datasets, place a significant burden on healthcare systems and ophthalmologists.
“The human population is steadily growing with increased life expectancy, impacting the prevalence of age-dependent diseases, including age-related macular degeneration (AMD). Health care systems are confronted with an increasing burden with rising patient numbers accompanied by ongoing developments of therapeutic approaches.”

AI as a Solution for Efficient and Objective AMD Assessment:

AI offers a fast and objective solution for assessing AMD across all disease stages.
Reliable and validated AI algorithms can help manage the growing patient numbers and maximize the benefits of multimodal imaging.
AI-based algorithms provide “reliable and precise quantification of retinal features in a fraction of the time a human examiner would need, providing more time for doctor-to- patient interactions.”
Automated analysis of data is becoming “key in the future management of AMD.”
AI models are capable of image recognition, classification, segmentation, and even predicting future progression.
Further benefits include “excellent reproducibility and the performance of some algorithms can even exceed the performance of human graders.”

Types of AI Applied in AMD:

Machine Learning (ML): Algorithms learn from human-expert-defined biomarkers or features (supervised learning). An example is training a classifier to recognize drusen for early AMD diagnosis.
Deep Learning (DL): Uses deep neural networks (DNNs) with multiple layers. Unlike ML, DL algorithms can learn to extract features and classify simultaneously (unsupervised or end-to-end learning). Performance improves with dataset size.
Convolutional Neural Networks (CNNs): A specific type of DNN particularly suitable for imaging data. Used for classification and pixel-wise segmentation and quantification of features like fluid, drusen, and retinal layers.
U-Net: A convolutional network adapted for tasks like retinal layer segmentation.
Transfer Learning: Enables algorithm development with small datasets by fine-tuning pre-existing CNNs trained on large datasets.
Recurrent Neural Networks (RNNs): Well-suited for sequence labeling and prediction tasks by considering past input.
Generative Adversarial Networks (GANs): Consist of a generator and discriminator working together to synthesize high-quality data, which can aid in data augmentation for training on imbalanced datasets and improving image quality (denoising, super-resolution).

AI Applications in Early and Intermediate AMD (iAMD):

Early stages are characterized by drusen and pigmentary alterations, often with limited symptoms.
OCT is becoming the key imaging modality for early detection and monitoring due to its detailed 3D visualization of features like drusen.
AI algorithms are being developed for automated classification of AMD on various imaging modalities (CFP, FAF, NIR, OCT).
Assessment of iAMD progression is challenging due to subtle changes and lack of validated clinical endpoints for trials.
Several OCT-based biomarkers for conversion to GA have been identified, including nascent GA, incomplete RPE and outer retinal atrophy (iRORA), wedge-shaped subretinal hyporeflectivity, RPE attenuation/disruption, drusenoid RPE detachment, RPE thickening, hyperreflective foci (HRF), subretinal drusenoid deposits (SDD), outer plexiform layer (OPL) subsidence, outer retinal layer thickness reduction, photoreceptor (PR) atrophy, hyporeflective cores in drusen, and high drusen volume.
ML and DL algorithms are used for automated layer segmentation and drusen volume quantification, which is associated with AMD progression.
Automated quantification of outer retinal layer integrity, such as Ellipsoid Zone (EZ) and Outer Nuclear Layer (ONL) thickness, is being explored as a quantifiable subclinical biomarker.
HRF and SDD quantification using AI are also being investigated as potential biomarkers for disease progression.

AI Prediction of Conversion to Advanced AMD:

Predicting progression is crucial for preventing vision loss and optimizing monitoring intervals.
Imaging provides multiple biomarkers preceding conversion to advanced AMD, including drusen dynamics (size change, fusion, regression).
ML models are being developed to predict progression based on OCT features, demographics, and potentially genetic profiles.
One ML model achieved an AUC of 0.75 for prediction within the first two years based on outer retinal layer segmentations and HRF.
Longitudinal OCT analysis identified ONL and EZ thinning associated with atrophy development and choroidal changes linked to nAMD onset.
Large datasets (like the PINNACLE consortium) are being used to develop algorithms predicting disease progression using supervised and unsupervised ML, and DL.
Innovative self-supervised learning methods are being explored to simulate morphological changes and provide insight into progression risk.
Predictive models can help clinicians estimate appropriate check-up times for early intervention.

AI Applications in Geographic Atrophy (GA):

GA is a progressive degeneration leading to irreversible vision loss.
FAF was the main imaging modality for demarcating atrophic areas, but OCT is now the gold standard for diagnosis and management.
AI-based algorithms provide accurate results comparable to human performance in segmenting GA on FAF and NIR.
Semi-automated and fully automated DL methods (like U-NET) are being developed for GA segmentation on FAF.
The equivalent biomarker for progression speed on OCT is RPE loss, visualized directly or through hypertransmission.
Automated segmentation of hypertransmission on OCT using DL is an accurate and reproducible method for longitudinal GA assessment.
Loss of the EZ on OCT is recognized as an important biomarker. Validated algorithms segmenting EZ and IZ are being established.
DL algorithms have demonstrated a significant protective treatment effect on RPE and photoreceptor integrity loss in clinical trials.

AI Prediction of GA Progression:

Observing longitudinal GA growth is important for patient management and assessing treatment efficacy.
AI allows automated quantification on OCT for both RPE and photoreceptors.
An association between GA growth rates and the ratio of photoreceptor integrity loss to RPE loss has been found.
Validated DL algorithms predict topographic progression of GA by analyzing RPE loss, photoreceptor integrity, and HRF in a spatio-temporal manner.
Higher progression rates have been associated with lesions closer to the fovea, HRF at the junctional zone, and thinner photoreceptor layers.
Multitask approaches using multimodal imaging (FAF and OCT) are being used to estimate GA lesion size and growth rates simultaneously.
These tools support clinicians in identifying progression patterns and speed for patient selection for treatment.

AI Applications in Neovascular AMD (nAMD):

nAMD is characterized by MNV proliferation and exudation of fluid (IRF, SRF, PED).
Anti-VEGF therapy is the gold standard treatment, but outcomes can be suboptimal in real-world practice due to undertreatment and development of atrophy/fibrosis.
Treatment regimens (PRN, treat-and-extend) depend on OCT-based monitoring of disease activity, generally defined as fluid recurrence or persistence.
Objective quantification of fluid compartments in OCT volumes is of high clinical relevance, but manual annotation is not feasible.
ML and DL algorithms are being developed for automated fluid volume quantification.
Pioneering work demonstrated that DL algorithms (like the Fluid Monitor) are suitable for automated fluid volume quantification in SD-OCT scans and have been validated on large datasets.
Automated fluid quantification is also being explored for home-based monitoring with Home OCT devices.

AI for Quantification of Disease Activity and Correlation to Function in nAMD:

Automated quantification of fluid volume allows for correlation with visual outcomes.
Post-hoc analyses of clinical trials and real-world data confirm that IRF, particularly in the central mm, is associated with worse visual outcomes.
High fluid volume at baseline and after loading doses can predict the need for more treatment during follow-up.
Automated quantification of other features like subretinal hyperreflective material (SHRM) and EZ attenuation has also shown associations with visual functional outcome.

AI Prediction of nAMD Progression and Treatment Outcomes:

Predicting treatment response and need is gaining clinical relevance due to interindividual variability.
ML algorithms can predict visual acuity from baseline OCT volumes, with baseline visual acuity and central IRF being important prognostic biomarkers.
Adding treatment response parameters to the model improves predictive value.
DL models are being developed to predict treatment requirements during follow-up.
AI algorithms are used to predict development of macular atrophy and fibrosis in real-world cohorts.
GANs are being explored for synthetic post-therapeutic OCT image reconstruction to potentially predict treatment outcomes.

Limitations of AI in AMD:

Need for high-quality and diverse datasets for algorithm training and validation to represent the real population.
Challenges with interpretability of AI decisions, particularly in “blackbox” models.
Algorithms need to adapt to varying imaging protocols and real-world image quality.
Lack of universally applicable algorithms due to novel imaging modalities and devices.
Concerns regarding patient privacy and data security when using cloud-based processing.
Financial burden of security systems, data storage, computational power, and commercially available algorithms.

Future Directions and Conclusion:

AI has entered clinical practice for AMD diagnosis and management.
AI is crucial for detecting and quantifying subclinical changes and evaluating OCT data beyond human capacity.
AI can reproduce learned information and outperform clinicians in terms of time savings and accessibility of knowledge.
Ongoing validation of AI models, including external and diverse datasets, is essential.
AI models for disease screening and risk stratification are expected to gain momentum and become routine.
The synergy between large OCT data and AI analysis will remain an essential part of ophthalmology practice, guiding personalized and targeted medicine.