AUTOMATIC DETECTION OF FOVEA AND OPTIC DISC USING DEEP NEURAL NETWORKS

Al-Bander, B.; Al-Nuaimy, W.; Parry, D.G.; Leach, S.; Zheng, Y.

Abstract

Design: This is a software development and evaluation study involving colour fundus images of the retina from people with diabetes.
Purpose: To investigate the feasibility of deep learning techniques to simultaneously detect the centres of the fovea and the optic disc (OD) from colour fundus images.
Methods: 1200 macula-centred digital colour fundus photographs from the publically available MESSIDOR dataset were used. The centres of the fovea and the OD in each image were marked up by expert graders (DGP and SL) as ground truth. 975 images and their annotated locations of the foveal and OD centres were used to train convolutional neural networks (CNNs) and the rest 225 images were used to evaluate the performance of the trained CNNs. In the pre-processing step, the intensity of each image was scaled between [0, 1] and the centres locations were scaled between [-1, 1]. The proposed CNNs comprises multiple convolution layers, max pooling layers, fully connected layers, dropout layers and output layer. Following the literature, an automatically detected foveal (resp. OD) centre is considered to be correct if the distance between it and the annotated foveal (resp. OD) centre is less than 0.5 OD diameter (ODD). The ODD was estimated by dividing the distance between the annotated foveal centre and the OD centre by 2.5).
Results: The accuracy is 95.1% for the detection of the foveal centre while 96.0% for the detection of the OD centre. The mean ± standard deviation (std) distance of the detected foveal centre from the annotated foveal centre is 0.151 ± 0.225 ODD. The distance of the detected OD centre from the annotated OD centre is 0.146 ± 0.139 ODD.
Conclusions: The proposed CNNs approach showed very promising results for simultaneously automated detection of the centres of the fovea and OD. Further optimisations on CNNs are underway for its introduction into general clinic practice.

Citation

Al-Bander, B., Al-Nuaimy, W., Parry, D., Leach, S., & Zheng, Y. (2016, June). AUTOMATIC DETECTION OF FOVEA AND OPTIC DISC USING DEEP NEURAL NETWORKS. Poster presented at 26th Meeting of the European Association for the Study of Diabetes Eye Complications Study Group (EASDec), Manchester, UK

• Presentation Conference Type: Poster
• Conference Name: 26th Meeting of the European Association for the Study of Diabetes Eye Complications Study Group (EASDec)
• Conference Location: Manchester, UK
• Start Date: Jun 23, 2016
• End Date: Jun 25, 2016
• Deposit Date: Nov 13, 2023
• Publisher URL: https://journals.sagepub.com/doi/full/10.5301/ejo.5000799