Identifying Peripheral Neuropathy in Colour Fundus Photographs Based on Deep Learning

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Diabetic neuropathy (DN) is a common complication of diabetes, affecting nerve function.
• Early detection of DN is crucial for preventing severe complications.
• Retinal imaging offers a non-invasive method for assessing systemic health in diabetic patients.

Purpose of the Study:

• To develop and validate a deep learning (DL) system for detecting diabetic neuropathy (DN) using retinal color images.
• To assess the performance of DL models in identifying DN in individuals with diabetes.
• To explore the utility of retinal imaging for DN screening in conjunction with diabetic retinopathy (DR) screening.

Main Methods:

• A dataset of 1561 retinal images from diabetic patients was used.
• Deep neural networks (Squeezenet, Inception, Densenet) were trained and validated.
• Performance was evaluated using Area Under the ROC Curve (AUC), with models tested on all images, images without DR, and images with DR.

Main Results:

• The DL system achieved an AUC of 0.8013 on the validation set and 0.7097 on the test set for predicting DN.
• Performance improved to an AUC of 0.8673 when analyzing images from patients with diabetic retinopathy (DR).
• The study demonstrated the potential of retinal images to identify individuals with DN.

Conclusions:

• Deep learning analysis of retinal images can effectively detect diabetic neuropathy (DN).
• Retinal imaging provides a valuable opportunity for early DN detection during routine diabetic retinopathy (DR) screening.
• This technology can facilitate patient education regarding their DN status.