Segmentation of retinal layers on OCT scans using deep learning

Abstract

This past decade, the use of artificial intelligence, and more precisely deep learning, has been really efficient in image processing to obtain good performances in object detection and image segmentation. In the medical field, retinal imaging represents an important area of research with a great clinical interest. Indeed, the observation of the retinal layers is helpful in the diagnosis, treatment and monitoring of plenty of retinal pathologies. In this context, this project was focused on using deep learning for OCT retinal layers segmentation. To do so, a UNET-VGG16 model has been employed and the method was evaluated on a Duke OCT database of 4780 B-scans. It succeeded in segmenting three retinal layers with an IoU of 0.529 and a Dice coefficient of 0.685. To go further, the use of data augmentation, pre-processing and post processing functions could solve some issues and improve the method.