A Non-Invasive Deep Learning Approach for Early Detection of Peripheral Arterial Disease–Diabetic Foot Ulcers Using Retinal Imaging: A Prospective Cohort Study

Peripheral Arterial Disease (PAD) is one of many serious vascular complications of diabetes mellitus, and it is a major contributor to diabetic foot ulcers (DFUs) and amputations. The importance of early detection is paramount; however, the existing diagnostic tools, the Doppler ultrasound and Ankle Brachial Index (ABI) testing, are expensive and not available in many countries. This article introduces an artificial intelligence (AI)-based project of detecting PAD non-invasively and predicting DFU risk through retinal fundus images. An adapted EfficientNet-B3 convolutional network with deep retina features combined with vessel density, fractal dimension, hemodynamic index, and vascular descriptors was created and tested on a clinically validated cohort of 1,928 diabetic patients with gold-standard PAD confirmation via ABI and Doppler ultrasound. With a traditional convolutional neural network and ResNet-50 model, the model has an AUC-Receiver Operating Characteristic of 96.1%, sensitivity of 94.2, and specificity of 91.5, and was superior to the models (P < 0.001). The biologically relevant vascular attention was validated by Grad-CAM visualization. This model emphasizes the opportunities of retinal imaging to be a cost-efficient, AI opportunistic PAD and DFU screening tool in the routine diabetic eye examination.