Optic disc (OD) is one of the most important parts of the retina. Accurate segmentation of the OD provides essential information about the health of the retina and aids in the diagnosis of various diseases that affect the retina, such as glaucoma and diabetic retinopathy (DR). Segmenting the OD manually can be influenced by numerous variables like image clarity and lighting, demanding expertise for precise boundary detection. The nature of this task emphasizes the need for accessible automated or semi-automated OD segmentation solutions to address the shortage of specialists in diagnosing retinal conditions effectively almost worldwide. In this study, an automatic method for OD segmentation in retinal images using a convolutional neural network (CNN) architecture, known as U-Net, was introduced. To enhance the U-Net's efficacy compared to previous studies, we adopted a two-step strategy: first, by cropping the image around the OD using the bounding box technique, and then by resizing it to 128×128 pixels. This method not only preserved robust OD segmentation but also significantly reduced computational time remarkably. Then, these images were enhanced using the contrast limited adaptive histogram equalization (CLAHE) to eliminate the noise and improve their qualities. After that, a U-Net model was constructed and trained to obtain segmented images. The proposed model was trained and evaluated using the public dataset ORIGA, and the predicted results were compared with the ground truth (GT) images. This method competed with other studies and achieved competing results, average value of each of the following metrics came to be: recall 95.33%, F-score 96.54%, and IoU 93.41%.

Title

IEEE Xplore

Publisher

IEEE

Publisher Country

Jordan

Publication Type

Prtinted only

Volume

--

Year

2025

Pages

--