Explicit Inclusion of Diabetes Mellitus Without Retinopathy Within Diabetic Retinopathy Prediction

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early detection of DR is crucial for preventing vision impairment.
• Current diagnostic methods may not sufficiently capture early retinal changes in diabetes mellitus (DM) without DR.

Purpose of the Study:

• To evaluate if explicitly modeling DM without DR as a distinct stage improves deep learning (DL) detection of early retinal changes.
• To enhance early risk identification for the DR severity spectrum.
• To assess the impact of specific retinal regions on DL classification performance.

Main Methods:

• Developed three DL classification models with varying granularity (3, 4, and 6 classes) incorporating DM without DR as a distinct stage.
• Utilized 6069 color fundus images from diverse DM and DR categories.
• Developed segmentation models for optic nerve head (ONH) and retinal vessels to analyze feature importance and spatial changes.

Main Results:

• The models demonstrated varying performance across stages, with improved discrimination for early DM without DR.
• Area under the curve (AUC) for DM without DR ranged from 65.7% to 80.3% across models.
• Perturbations in retinal vessels significantly impacted performance (16-31%), and increased DR severity correlated with saliency-to-ONH distance (r=0.69-0.72).

Conclusions:

• Explicitly modeling diabetes without retinopathy enhances early-stage discrimination and reveals feature-reliance shifts.
• Vessel and saliency-based analyses identified subtle retinal changes preceding clinical DR.
• This approach aids the development of clinically useful artificial intelligence (AI) tools for DR risk identification.