A Deep Learning Framework for Identifying Zone I in RetCam Images

Abstract

Retinopathy of prematurity (ROP) has been one of the worldwide causes of blindness among children. Grading and treatment guidelines of ROP are mainly based on zone, stage, and plus disease. For serious ROP, identifying zone is more important than staging. However, identifying zone I from RetCam fundus images is not accurate and subjective by ophthalmologists. To address it, we develop a new deep learning framework to automatically identify zone I from RetCam images. Specifically, we train a deep convolutional neural network (DCNN) algorithm based on the RetCam images. The disc and macular center in terms of the threshold of intersection over union (IOU) were identified automatically. The algorithm is validated on fundus images and results show that zone I identification accuracy of 91% is achieved when the IOU threshold is 0.8. The obtained promising identification accuracy of zone I from the RetCam images indicates the potential applications in ROP grading, monitoring, and prognosis for infants.

Keywords

• Retinopathy of prematurity
• deep convolutional neural network
• optic disc
• macular center
• zone I