June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Deep Binary Descriptor-based Identification of Retinal Lesions for Diabetic retinopathy
Author Affiliations & Notes
  • Jian Lian
    School of Intelligence Engineering, Shandong Management University, Jinan, Shandong, China
  • Xinbo Yang
    Shandong Normal University, Jinan, Shandong, China
  • Footnotes
    Commercial Relationships   Jian Lian None; Xinbo Yang None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 1740 – F0200. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jian Lian, Xinbo Yang; Deep Binary Descriptor-based Identification of Retinal Lesions for Diabetic retinopathy. Invest. Ophthalmol. Vis. Sci. 2022;63(7):1740 – F0200.

      Download citation file:

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

  • Supplements

Purpose : The detection and classification of retinal lesions play a vital role in screening diabetic retinopathy, one primary cause of visual impairment globally. Extensive research on deep learning has shown their good performance in this area. However, these methods suffer from the image samples’ scarcity and diversity of morphological characteristics. This study aims to establish a binary descriptor-based classification instrument for discriminating the retinopathy slices by leveraging deep backbone models capable of resolving complex cases.

Methods : 4 datasets of retinal slices were collected in this study. In total, there are three types of retinal images used as the input for the proposed deep learning-based pipeline, which consists of 3 phases (as shown in Fig. 1):
We added affine transformations to the raw samples.
What fed the original and deformed retinal images into one backbone model.
The generated feature for each slice was fed into a non-linear classifier to discriminate three categories of diabetic cases.
Note that three constraints, including affine transformation in variation, even distributed, and quantization loss minimization was imposed on the features rendered as a binary descriptor. The structure of the proposed approach is shown in Figure 1.

Results : The proposed framework is superior to the state-of-the-art algorithms in accuracy, sensitivity, and precision. The ten-fold cross-validation accuracy of the presented approach on the samples is 98.35%.

Conclusions : The findings indicate that the proposed approach is a potentially helpful instrument for retinal lesion identification. By constraining the output features extracted from the input images it not only can increase the generalization of the deep learning models but also guarantee accuracy.

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


The framework of the presented method and the general outcome.

The framework of the presented method and the general outcome.


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.