Quaternion Neural Network on Detecting Diabetic Retinopathy Using Fundus Images

Diabetic retinopathy (DR) is an eye disease due to the effects of diabetes on the eye, and its damage on the nerve tissue in the retina. DR involves vascular malformations that is strongly correlated with the duration of diabetes and is the leading cause of blindness in adults. The disease presents various lesions, the most significant of which are microaneurysms, exudates, and retinal vessel malformations. Retinal images assist physicians diagnosing and monitoring DR progression, while image processing and machine vision methods enhance the speed and accuracy of diagnosis. A review of existing DR diagnostic methods reveals that feature extraction from retinal images followed by classification using a convolutional neural network yields the best results. This study proposes a diagnostic framework that employs a quaternion convolutional neural network to improve DR detection accuracy. The dataset used, named Al-Zahra, consists of ۳,۸۷۲ retinal fundus images captured with a Canon CX-۱ device at Al-Zahra Ophthalmology Hospital in Zahedan. Simulation results on the Al-Zahra database demonstrate high detection accuracy: ۹۸.۸۷% for retinal vessel malformations, ۹۸.۳۹% for exudates, and ۹۸.۱۳% for microaneurysms.

کلیدواژه ها:

diabetic retinopathy ، microaneurysm ، exudate ، quaternion neural network ، Statistical features