A Deep Learning Approach to Improve Retinal Structural Predictions and Aid Glaucoma Neuroprotective Clinical Trial...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Accurate measurement of retinal nerve fiber layer (RNFL) and ganglion cell-inner plexiform layer (GCIPL) thickness is crucial for diagnosing and monitoring glaucoma.
• Clinical trials for neuroprotection therapies require robust outcome measures and efficient trial designs.

Purpose of the Study:

• To evaluate a deep learning (DL) regression model for predicting macula GCIPL and optic nerve head (ONH) RNFL thickness.
• To assess the utility of DL-based thickness predictions in glaucoma neuroprotection clinical trials.

Main Methods:

• A cross-sectional study utilizing data from 1096 eyes of 550 glaucoma patients.
• Spectralis imaging data (GCIPL and RNFL scans) were used to train and test DL models.
• Models predicted hemiretinal GCIPL and RNFL thickness using data from the fellow eye and one hemiretina of the probe eye.

Main Results:

• The DL model achieved high accuracy in predicting GCIPL thickness (global r² of 0.90 and 0.86).
• RNFL thickness prediction showed slightly lower but still strong performance (global r² of 0.75 and 0.84).
• Model performance remained robust even in eyes with more severe glaucoma.
• Simulations indicated an 11-fold reduction in patient numbers for clinical trials using DL-based predictions.

Conclusions:

• Deep learning models can accurately estimate GCIPL and RNFL thickness.
• DL-based internal controls can significantly reduce sample size requirements for glaucoma clinical trials.
• This approach facilitates the investigation of novel glaucoma neuroprotection therapies.