Diabetic peripheral neuropathy identification using enface optical coherence tomography and multi-head attention deep learning algorithm.

Objective: Diabetic peripheral neuropathy (DPN) is a common chronic complication of diabetes, but current diagnostic methods are limited by invasiveness, poor sensitivity, or subjectivity. This study aims to develop a non-invasive, reliable diagnostic tool using multimodal optical coherence tomography (OCT) images and a deep learning (DL) algorithm with multi-head attention for early DPN detection. Methods: A multi-head attention-based DL model was constructed, with ResNet-18 as the feature extractor to fuse and classify enface OCT images from different retinal layers. A total of 3264 OCT images from 544 eyes of 435 diabetic patients were enrolled. The model was evaluated via fivefold cross-validation on the training dataset (Dataset A, n = 267) and further validated on a temporal validation dataset (Dataset B, n = 168). Single-layer contrast experiments were conducted to identify the most predictive retinal layer, and Gradient-weighted Class Activation Mapping (Grad-CAM) was used for model visualization. Results: The proposed model achieved an average area under the curve (AUC) of 0.719 in fivefold cross-validation and an AUC of 0.721 in the temporal validation dataset. Among all retinal layers, the avascular layer provided the highest predictive value for DPN (average AUC = 0.707), with significant differences in performance compared to other layers ( Conclusion: The multi-head attention-based DL model effectively identifies DPN using non-invasive OCT images, with the avascular layer providing critical information. This approach provides a promising clinically feasible early screening strategy, and photoreceptor defects may serve as a potential DPN biomarker, requiring further validation.