Learning to Correct Axial Motion in Oct for 3D Retinal Imaging

Abstract

Optical Coherence Tomography (OCT) is a powerful technique for non-invasive 3D imaging of biological tissues at high resolution that has revolutionized retinal imaging. A major challenge in OCT imaging is the motion artifacts introduced by involuntary eye movements. In this paper, we propose a convolutional neural network that learns to correct axial motion in OCT based on a single volumetric scan. The proposed method is able to correct large motion, while preserving the overall curvature of the retina. The experimental results show significant improvements in visual quality as well as overall error compared to the conventional methods in both normal and disease cases.

Joint Motion Correction and 3D Segmentation with Graph-Assisted Neural Networks for Retinal OCT

“…The first 55 OCT volumes are used for training, and next 5 volumes are used for validation, and the last 55 volumes are used for testing. Since the DME dataset [7] has been motion corrected, we include simulated motion on the input OCT volumes to test the performance of our proposed motion correction approach as described in [21].…”

“…1. In the first stage, motion artifacts in the input OCT volume V corrected using the motion correction network proposed in [21] based on the 3D input volume and a rough 3-class segmentation of the entire retina using a 2D segmentation method. In the second stage, the detailed retinal layers S pred are classified using a segmentation network with 3D input based on the motion-corrected OCT volume.…”

“…In the second stage, the detailed retinal layers S pred are classified using a segmentation network with 3D input based on the motion-corrected OCT volume. The motion correction network [21] is designed to predict a 2D displacement map D to remove axial motion artifacts from 3D OCT volume, and the top (inner limiting membrane, ILM) and bottom (Bruch's membrane, BM) segmentation boundaries of the retina where BM and ILM layers improve correction performance. The motion correction network is based on a modified U-Net with residual blocks, which can be applied to different resolutions.…”

“…The motion correction network is based on a modified U-Net with residual blocks, which can be applied to different resolutions. More information can be found in [21]. The OCT volume is then warped based on the predicted the 2D displacement map.…”

“…Besides the standard data augmentation techniques such as random cropping and horizontal flipping, we also include simulated shearing along the X axis to add another degree of freedom. The method to generate simulated shearing is described in [21].…”