LONGL-Net: temporal correlation structure guided deep learning model to predict longitudinal age-related macular...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Age-related macular degeneration (AMD) is a leading cause of blindness globally, with increasing prevalence.
• Current methods for AMD assessment rely on manual grading of Color Fundus Photographs (CFP), which can be subjective and time-consuming.
• Predicting the longitudinal progression of AMD is crucial for timely clinical intervention but remains challenging.

Purpose of the Study:

• To develop and validate a deep learning model for simultaneous grading of current AMD severity and prediction of future late-AMD progression using CFP.
• To introduce a novel temporal-correlation-structure-guided Generative Adversarial Network (GAN) for analyzing longitudinal CFP data.
• To provide interpretable predictions by forecasting AMD symptoms in future retinal images.

Main Methods:

• A deep learning classification model, LONGL-Net, was designed incorporating a temporal-correlation-structure-guided GAN.
• The model was trained and evaluated on approximately 30,000 CFP images from 4,628 participants in the Age-Related Eye Disease Study (AREDS).
• External validation was performed on 300 CFP images from the UK Biobank dataset.

Main Results:

• The LONGL-Net model achieved an average Area Under the Curve (AUC) of 0.905 and an accuracy of 0.762 for the 3-class classification task (simultaneous grading and future prediction) on the AREDS dataset.
• On the UK Biobank dataset, the model demonstrated an average accuracy of 0.905 and a sensitivity of 0.797 for grading CFP images.
• The GAN component provided interpretability by forecasting future AMD symptoms.

Conclusions:

• The proposed LONGL-Net model effectively grades current AMD severity and predicts future disease progression from CFP.
• This deep learning approach offers a promising tool for early identification of individuals at risk of late-AMD, enabling proactive clinical management.
• The model's ability to analyze temporal changes in retinal images enhances its clinical utility for personalized AMD patient care.