Evaluation of a portable retinograph assisted by convolutional neural networks for early diagnosis and grading of diabetic retinopathy

Abstract

Diabetic retinopathy (DR) is one of the leading causes of vision loss worldwide, especially in regions with limited access to ophthalmological services. The aim of this study was to develop and evaluate a low-cost portable retinograph assisted by Deep Learning for early detection and severity classification of DR. The system integrates a portable device based on indirect ophthalmoscopy, 3D printing, and smartphone image acquisition, together with convolutional neural networks (CNNs) for automated analysis. A combined dataset from the Messidor, APTOS, and EyePACS databases, comprising approximately 49,430 retinal images, was used for training and validation. Image preprocessing, class balancing, and transfer learning techniques were applied using MobileNetV2 and DenseNet121 architectures, evaluated in binary and multiclass classification tasks. The results demonstrated that the prototype captured fundus images with sufficient quality for clinical visualization. In conclusion, the proposed system represents a feasible and accessible solution for DR screening in telemedicine and resource-limited settings.