Deep Learning Based Retinal Blood Vessel Segmentation of Multiple Optical Coherence Tomography En-Face Images in Cases of Optic Disc Swelling

Abstract

Purpose : En-face projection images of the retinal pigment epithelium (RPE) are frequently used for optical-coherence-tomography-based segmentation of retinal vessels. However, in cases of optic disc swelling, retinal vessel visibility directly under the swollen regions in the center of RPE en-face projection images is often poor. Motivated by our observation that in cases of optic disc swelling, the inner and total retina en-face images provide better vessel visibility of the optic disc central region and the RPE en-face image provides better visibility of outer regions (shown in Fig 1(b)), in this work, we propose a deep-learning approach (using a U-Net architecture) to provide a vessel probability map by simultaneously considering three en-face projection images as input.

Methods : A 3D graph-based algorithm was used to segment the retinal layers and the three 2D en-face images (shown in Fig 1(b)) were generated. We performed leave-one-out cross-validation on a dataset of 18 subjects having optic disc swelling. The deep neural network provides an output vessel probability map. A human expert vessel tracing by combining information from en-face images of RPE, inner-retinal, total retinal and registered fundus image served as the reference standard. The vessels from the RPE en-face image only were also manually traced to compare with the performance of the proposed approach.

Results : Three quantitative measurements: 1) AUC (area under the receiver operating characteristic curve), 2) coefficient of determination (R²) and 3) mean squared error (MSE) were used to evaluate the method. The deep-learning-based approach had an AUC 0.94, mean R² of 0.35, and mean MSE of 0.049 while the manual tracing from the RPE en-face image only had an AUC of 0.82, mean R² of 0.52, and mean MSE of 0.066.

Conclusions : The proposed method provides promising performance to segment the retinal vessels obscured by image shadows in cases of optic disc swelling compared to manual tracing. The proposed approach can be extended to perform 3D vessel segmentation in cases of optic disc swelling.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Fig 1: (a) OCT central B-scan with automated 3D graph-based layer segmentation. (b) Visibility of vessels in different levels of optic disc swelling. The blue arrows indicate the visibility changes.

 

Fig 2: Example of vessel segmentation in four subjects with various levels of optic disc swelling.