June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Exploring Deep-Learning on SD-OCT images as an Effective Mitigation Technique for Hydroxychloroquine-Induced Retinal Toxicity
Author Affiliations & Notes
  • Harsh Bandhey
    Electrical and Computer Engineering, Duke University, Durham, North Carolina, United States
  • Amanda Del Risco
    School of Medicine, Duke University School of Medicine, Durham, North Carolina, United States
  • Qitong Gao
    Electrical and Computer Engineering, Duke University, Durham, North Carolina, United States
  • Jay Rathinavelu
    School of Medicine, Duke University School of Medicine, Durham, North Carolina, United States
  • Miroslav Pajic
    Electrical and Computer Engineering, Duke University, Durham, North Carolina, United States
  • Majda Hadziahmetovic
    Department of Ophthalmology, Duke University, Durham, North Carolina, United States
  • Footnotes
    Commercial Relationships   Harsh Bandhey None; Amanda Del Risco None; Qitong Gao None; Jay Rathinavelu None; Miroslav Pajic None; Majda Hadziahmetovic None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 3330 – F0139. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Harsh Bandhey, Amanda Del Risco, Qitong Gao, Jay Rathinavelu, Miroslav Pajic, Majda Hadziahmetovic; Exploring Deep-Learning on SD-OCT images as an Effective Mitigation Technique for Hydroxychloroquine-Induced Retinal Toxicity. Invest. Ophthalmol. Vis. Sci. 2022;63(7):3330 – F0139.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : There is no precise method of individualized risk assessment and automated screening for macular retinopathy due to hydroxychloroquine (HCQ) induced retinal toxicity. We propose the explorative analysis of deep learning methods on spectral-domain optical coherence tomography (SD-OCT) to detect retinopathy before topographical fundus damage and further predict individualized SD-OCT-based risk assessment for HCQ toxicity.

Methods : Clinical Data from 73 patients on HCQ treatment, including SD-OCT, multifocal electroretinogram (mfERG), and autofluorescence (FAF) images, were collected and classified as normal vs. abnormal according to overall clinical assessment and mfERG results (as per Figure 1).
For detection of HCQ induced retinopathy, an Inception-v3 model was trained on augmented images from corresponding “normal” and “abnormal” SD-OCT images using transfer learning on ImageNet pre-trained weights. Performance improvement on Image preprocessing techniques such as Contrast Limited Adaptive Histogram Equalization (CLAHE) and Block Matching and 3-D (BM3D) Filtering were also analyzed.
For prediction of risk of HCQ toxicity, a Long Short Term Memory-based Classifier is proposed with the same structured data as shown below in Figure 2.

Results : For detection, on preprocessed images augmented from 228 “abnormal” SD-OCTs and 128 “normal” SD-OCTs with a 80-20 train-test split, the Inception-v3 model gave a precision of 0.72, recall of 0.92, F1 score of 0.81 and an Accuracy of 0.81 for detecting “abnormal” conditions.
Contrast Limited Adaptive Histogram Equalization (CLAHE) significantly improved the performance. However, Block Matching and 3D Filtering (BM3D) was too slow to be performed on an augmented dataset. For prediction, we had only three patients with advanced “bull’s eye” maculopathy presented over a period of time; thus, prediction analysis could not be performed.

Conclusions : Explorative analysis of deep-learning methods on SD-OCT shows that effective screening methods could be developed for early macular retinopathy detection due to HCQ toxicity with proper Image preprocessing. DL-based HCQ toxicity risk predictors require a larger dataset, and further analysis is required for this task.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1. Labeling Methodology

Figure 1. Labeling Methodology

 

Figure 2. Proposed Prediction Architecture

Figure 2. Proposed Prediction Architecture

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×