Detecting Double Layer Signs with OCT volumes using a 3D Convolutional Neural Network (CNN)

Abstract

Purpose : The presence of a double layer sign (DLS) on structural OCT B-scans is a critical predictor for subclinical choroidal neovascularization (CNV), a stage of non-exudative type 1 macular neovascularization (MNV) before onset of exudation. We sought to develop a 3D CNN to detect DLS of any size using only structural OCT B-scans.

Methods : Eyes with a DLS and eyes with drusen (Dr), to serve as a control, were imaged using 6x6mm swept source OCT angiography (SS-OCTA, PLEX Elite 9000, Carl Zeiss Meditec, Dublin, CA). Each scan pattern consisted of 500 A-scans per B-scan with each B-scan repeated twice at each of 500 B-scan positions along the 6mm y-axis. The OCTA data was used for manual labeling of DLS and Dr; only the structural OCT was used for deep learning.

Results : A total of 232 eyes (196 patients; 173 with DLS and 53 with Dr) were imaged using the SS-OCTA scan pattern. The deep learning model for multi-region segmentation (Figure 1) labels DLS and Dr on a single B-scan image (3D-2D model). We generated dense annotations by integrating manual annotations and predicted segmentation (Figure 2). After refining the labels, we trained a final 3D convolutional model that segments volumetrically (3D-3D model). Finally, eyes with MNV were identified based on en-face projection maps of the predicted masks. Accuracy of final classification was 92.85% (3D-2D model) and 94.28% (3D-3D model). Mean intersection over union (IoU) was DLS: 31.39%, Dr: 12.23% for the 3D-2D model, and DLS: 57.36%, Dr: 25.20% for the 3D-3D model.

Conclusions : Our network can detect DLS from structural B-scans alone by applying an annotation refinement technique for 3D CNN to a dataset with coarse annotations.

This is a 2021 ARVO Annual Meeting abstract.

 

A. Process overview. B. Annotation refinement process; we built a 3D-2D model, then generated prediction masks for each B-scan to obtain dense segmentation masks with pseudo labels. Blue: double layer sign; green: Dr. We expanded the original annotation using these interim results, then trained our 3D-3D model.

 

A & B. Original annotations/predicted masks and the en-face projection map. C. For vertical slice (a), we mapped original annotations (b) and extracted all lesions that overlapped with or were very close to the original annotations (c). The extracted lesions were relabeled (d) so that each connected lesion had a consistent label. D. The criteria in C were applied to the normal B-scan direction to obtain the refined annotation.