Abstract
Purpose :
Ophthalmologists examine the retinal vasculature of a patient in order to ascertain the condition of the patient's ocular health. A physician's examination of retinal blood vessels is a difficult and time-consuming task. This study aims to advance the development of machine learning and artificial intelligence technologies to aid ophthalmologists in making quicker and more precise patient diagnoses.
Methods :
Retinal images were obtained from the DRIVE dataset which is a publicly available dataset including forty retinal pictures for training algorithms to segment blood vessels. This dataset includes a total of 40 photos captured with a Canon CR5 non-mydriatic 3CCD camera with a 45-degree field of view. The model was evaluated using three metrics: sensitivity, specificity, and accuracy. As specified by the dataset, images from the DRIVE dataset were divided into train and validation sets. Batches of five photos were utilized to train the model with an initial learning rate of 1x10-3. A U-Net model was fine-tuned for a total of 50 epochs, with each epoch lasting around three seconds, for a total training duration of approximately one minute and thirty seconds.
Results :
Segmentation of retinal blood vessels was conducted using a U-Net model. This was a model of a convolutional network having encoder and decoder sections. The model was trained for 50 epochs, with each epoch's sensitivity, specificity, and accuracy reported. Upon completion of the model's training, it had a sensitivity of 87%, specificity of 97%, and accuracy of 97%. Our model showed the highest sensitivity, specificity, and accuracy when compared to previous research conducted on the DRIVE dataset.
Conclusions :
This study contributed to the construction of a U-Net model capable of performing effective blood vessel segmentation on DRIVE pictures. As a U-Net model, not only can the model discriminate between blood vessels and other elements of the eye, but it can also build a map consisting exclusively of blood vessels. In analyzing the structure and organization of blood vessels, it can be used to develop methods to identify and recognize pathology. With the aid of the technology created in this study, doctors may be able to recognize formations that point to a number of diseases, including retinopathy, diabetes, edema, and potentially macular degeneration.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.