Detecting Retinal Nerve Fiber Layer Segmentation Errors on Spectral Domain-Optical Coherence Tomography with a Deep Learning Algorithm

Abstract

Purpose : To propose a deep learning algorithm that detects errors in retinal never fiber layer (RNFL) segmentation from spectral-domain optical coherence tomography (SDOCT) B-scans.

Methods : A cross-sectional study with 25,250 OCT B-scans reviewed for segmentation errors from 1,363 eyes of 709 subjects. The sample was randomly divided into validation plus training (80%) and test (20%) sets. SDOCT B-scans that had been labeled for quality by human graders were used as the reference standard to train a deep learning convolutional neural network to detect RNFL segmentation errors. The performance of the deep learning algorithm was evaluated in the test sample by outputting a probability of having a segmentation error for each B-scan. The accuracy and the ability of the algorithm to detect segmentation errors was evaluated with the area under the receiver operating characteristic curve (ROC).

Results : Mean deep learning probabilities of segmentation error in the test sample were 0.91 ± 0.17 vs. 0.12 ± 0.21 (P<0.001) for scans with and without segmentation errors, respectively. The deep learning algorithm had an area under the ROC curve of 0.981 (95% CI: 0.975 to 0.986) with an overall accuracy of 92.8%.

Conclusions : We introduced a novel deep learning approach to assess RNFL segmentation errors in SDOCT B-scans. This algorithm can help clinicians and researchers review images for artifacts in SDOCT tests in a timely manner and avoid inaccurate diagnostic interpretations.

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

 

Figure 1. Spectral-domain optical coherence tomography (SDOCT) B-scans with segmentation errors correctly detected by both the reading center and the deep learning algorithm. Class activation maps (heatmaps) on the right show the regions of the B-scans that had greatest weight in the deep learning algorithm classification. It is possible to see that the deep learning algorithm identified (A) errors of segmentation involving both the delineation of the internal limiting membrane, (B) as well as the posterior boundary of the RNFL, (C) multiple errors in the same scan, and (D) even very small segmentation errors.