Deep Learning-Based Automatic Detection of Ellipsoid Zone Loss in Spectral-Domain OCT for Hydroxychloroquine Retinal Toxicity Screening

Abstract

Purpose: Retinal toxicity resulting from hydroxychloroquine use manifests photoreceptor loss and disruption of the ellipsoid zone (EZ) reflectivity band detectable on spectral-domain (SD) OCT imaging. This study investigated whether an automatic deep learning-based algorithm can detect and quantitate EZ loss on SD OCT images with an accuracy comparable with that of human annotations.Design: Retrospective analysis of data acquired in a prospective, single-center, case-control study.Participants: Eighty-five patients (168 eyes) who were long-term hydroxychloroquine users (average exposure time, 14 AE 7.2 years).Methods: A mask region-based convolutional neural network (M-RCNN) was implemented and trained on individual OCT B-scans. Scan-by-scan detections were aggregated to produce an en face map of EZ loss per 3dimensional SD OCT volume image. To improve the accuracy and robustness of the EZ loss map, a dual network architecture was proposed that learns to detect EZ loss in parallel using horizontal (horizontal mask region-based convolutional neural network [M-RCNN H ]) and vertical (vertical mask region-based convolutional neural network [M-RCNN V ]) B-scans independently. To quantify accuracy, 10-fold cross-validation was performed.Main Outcome Measures: Precision, recall, intersection over union (IOU), F1-score metrics, and measured total EZ loss area were compared against human grader annotations and with the determination of toxicity based on the recommended screening guidelines.Results: The combined projection network demonstrated the best overall performance: precision, 0.90 AE 0.09; recall, 0.88 AE 0.08; and F1 score, 0.89 AE 0.07. The combined model performed superiorly to the M-RCNN H only model (precision, 0.79 AE 0.17; recall, 0.96 AE 0.04; IOU, 0.78 AE 0.15; and F1 score, 0.86 AE 0.12) and M-RCNN V only model (precision, 0.71 AE 0.21; recall, 0.94 AE 0.06; IOU, 0.69 AE 0.21; and F1 score, 0.79 AE 0.16). The accuracy was comparable with the variability of human experts: precision, 0.85 AE 0.09; recall, 0.98 AE 0.01; IOU, 0.82 AE 0.12; and F1 score, 0.91 AE 0.06. Automatically generated en face EZ loss maps provide quantitative SD OCT metrics for accurate toxicity determination combined with other functional testing.Conclusions: The algorithm can provide a fast, objective, automatic method for measuring areas with EZ loss and can serve as a quantitative assistance tool to screen patients for the presence and extent of toxicity.

Automated Identification and Segmentation of Ellipsoid Zone At-Risk Using Deep Learning on SD-OCT for Predicting Progression in Dry AMD

“…Additionally, the current method retained this high performance across different device types (i.e., Cirrus and Spectralis devices), whereas the previous reports are limited to only one device type [ 33 , 34 ]. The pixel-wise segmentation accuracy of EZ At-Risk regions of 90% described in the current report is in line with the previous reports from other diseases [ 29 , 30 , 31 ]. Another important distinction from previously reported methods is the exclusion of regions with EZ loss in areas of pre-existing GA from the model training and analysis.…”

“…This pipeline had an accuracy of 85% with a sensitivity of 85% and specificity of 85% [ 28 ]. Another automatic DL-based quantification algorithm was used to detect EZ loss in 85 hydroxychloroquine retinopathy patients with an overall accuracy of 90% [ 29 ]. Similarly, automated EZ integrity loss detection in mild diabetic retinopathy achieved an accuracy of 90% in a small cohort of 13 patients [ 30 ].…”

Validation of a Deep Learning-Based Algorithm for Segmentation of the Ellipsoid Zone on Optical Coherence Tomography Images of an Ush2a-Related Retinal Degeneration Clinical Trial

“…Automatic algorithms have been developed to analyze the EZ on SD-OCT images from a variety of retinal diseases. [8][9][10][11][12][13][14][15][16][17][18][19][20] Although most of these algorithms were developed for a specific condition, generalizability is an important aspect of clinical applicability whereby the algorithm can be applied to other conditions. Trainable algorithms, such as modern deep learning-based algorithms, have especially strong potential to do so, given the appropriate training dataset.…”

Hydroxychloroquine Screening Guidelines: Best Evidence, Controversies, and Future Areas of Research

“…Studies have already demonstrated promising data on machinelearning models that incorporate clinical and/or structural quantitative data in detecting toxicity and predicting those at risk of progression. 43,44 Eventually, raw functional and structural data can be used, minimizing reliance on scan quality and leading to the identification of new imaging biomarkers for early toxicity which could provide novel insights into the mechanism of HCQ toxicity. '…”