Comparative performance of deep learning architectures for diabetic peripheral neuropathy detection using corneal...

This study aims to develop a deep learning algorithm (DLA) using the InceptionV3 architecture for effective diabetic peripheral neuropathy (DPN) screening via corneal confocal microscopy (CCM) images.

Retrospective study.

Ophthalmology Centre of General Hospital.

127 participants enrolled: 33 healthy participants, 57 diabetic patients with DPN (DPN+) and 37 diabetic patients without DPN (DPN-).

Not applicable.

The CCM image dataset, which was collected from participants (with five images per eye), was randomly divided into training, validation and test subsets in a 7:1:2 ratio. The images were preprocessed, augmented and used to train the InceptionV3 model. We compared its performance against the ResNet, DenseNet and Swin Transformer models. Performance was evaluated using accuracy, recall, F1 score and area under the curve (AUC) metrics.

For single-participant predictions, the InceptionV3 model achieved the highest accuracy (0.9231), recall (0.8846), F1 score (0.9020) and AUC (0.9534) compared with the other models. For single-image predictions in the three-class classification task of CCM images, the InceptionV3 model achieved a precision of 0.8385, a recall of 0.9083, an F1 score of 0.8720 and an AUC of 0.8769 for predicting DPN+.

The InceptionV3-based DLA model achieved superior performance compared with traditional convolutional neural network architectures like ResNet and DenseNet, and the Swin transformer model, highlighting its potential for effective DPN screening.