Detection of Diabetic Peripheral Neuropathy Using Enface OCT and Multi-Head Attention DL

Overview

This study developed a deep learning algorithm utilizing multi-head attention mechanisms to analyze enface optical coherence tomography (OCT) images for early detection of diabetic peripheral neuropathy (DPN). The approach demonstrated potential for non-invasive, sensitive identification of DPN in diabetic patients, addressing limitations of current diagnostic methods.

Background

Diabetic peripheral neuropathy (DPN) is a common and serious complication of diabetes, leading to sensory loss and increased risk of ulcers and amputations. Early detection is critical but current diagnostic tools like nerve conduction velocity tests and skin biopsies have limitations including invasiveness and low sensitivity. Optical coherence tomography (OCT) offers a non-invasive imaging technique that can assess retinal nerve fiber layers, which may serve as biomarkers for DPN. Deep learning methods, especially those incorporating multi-head attention mechanisms, can enhance image analysis by capturing complex features, potentially improving early DPN diagnosis.

Data Highlights

Participants were diabetic patients undergoing OCT angiography at Qilu Hospital between July 2023 and September 2024. Imaging used a 400 kHz swept-source OCTA device with 3.8 µm axial and 10 µm transverse resolution, scanning an 18 mm × 18 mm macular area. The study excluded patients with ocular or systemic conditions that could confound results. DPN diagnosis followed the Toronto Consensus Criteria requiring objective nerve function abnormalities and neuropathic symptoms or signs. Diabetes diagnosis adhered to 2024 ADA standards using fasting plasma glucose, oral glucose tolerance, HbA1c, or random plasma glucose thresholds.

Key Findings

• The proposed deep learning algorithm integrated multimodal enface OCT images to identify DPN with high sensitivity.
• Multi-head attention mechanisms enabled the model to capture both global and local retinal features relevant to neuropathy.
• Class activation maps highlighted specific retinal regions contributing to classification, aiding interpretability.
• The non-invasive OCT-based approach offers a clinically feasible alternative to invasive or subjective DPN diagnostic methods.
• The study population was rigorously selected to exclude confounding ocular and systemic diseases, enhancing result validity.

Clinical Implications

This automated OCT image analysis method can facilitate early, non-invasive detection of diabetic peripheral neuropathy, potentially enabling timely intervention to prevent progression. Incorporating multi-head attention deep learning models into clinical workflows may improve diagnostic accuracy and reduce reliance on invasive or subjective testing. This approach could be particularly valuable in screening high-risk diabetic populations.

Conclusion

The study demonstrates that multi-head attention deep learning applied to enface OCT images is a promising tool for early detection of diabetic peripheral neuropathy. This technique may enhance clinical management by providing a sensitive, non-invasive diagnostic option.

Related Resources & Content

1. Toronto Consensus Criteria 2010 -- Diabetic Peripheral Neuropathy Diagnosis
2. American Diabetes Association 2024 -- Diabetes Diagnosis Standards