Título: YOLO11 and YOLO12 Deep Learning Models for Diabetic Retinopathy Classification: A Comparative Analysis
Autores: Maria Cecília Alves Castro, Carlos Victor Gonçalves Moura, Débora Ferreira de Assis & Paulo César Cortez
Resumo: Diabetic retinopathy (DR) is a progressive microvascular complication of diabetes mellitus that affects the retina, causing changes in the blood vessels. In its early stages, the disease is asymptomatic, but as microvascular changes progress, the patient may experience impaired vision. The disease can be detected early through a fundus examination, in which the ophthalmologist looks for characteristic lesions in the image. Professionals often have difficulty in diagnosis due to the high volume of exams and the small size of the lesions in the image. DR can be prevented with glycemic control and regular eye exams, and early detection is essential to preserve the patient’s vision. Computer vision techniques can assist professionals in the diagnosis of DR through automatic image classification of the disease. This study aims to develop a multi-class classification method to identify the severity level of DR. To this end, the DDR and IDRiD dataset were used to train and validate the classification models. The YOLO11 and YOLO12 base models were trained in this task, and YOLO11 was also trained starting with pre-trained weights. A class reduction approach was also used, reducing the number of classes from 5 to 3 by merging classes from similar DR severity levels. The pre-trained YOLO11 model obtained the best results, achieving 71.04% mean accuracy for 5-class classification and 86.15% for 3-class. The model also achieved 88.33% ¨accuracy classifying the earlier stages of DR. Thus, the proposed method has the potential to aid ophthalmologists in the clinical diagnosis of DR, contributing an efficient and accessible approach to DR arge-scale screening.
Palavras-chave: diabetic retinopathy; image classification; YOLO; deep learning.
Páginas: 7
Código DOI: 10.21528/CBIC2025-1191909
Artigo em PDF: CBIC_2025_paper1191909.pdf
Arquivo BibTeX:
CBIC_2025_1191909.bib