Investigative Ophthalmology & Visual Science Cover Image for Volume 61, Issue 7
June 2020
Volume 61, Issue 7
Free
ARVO Annual Meeting Abstract  |   June 2020
Image quality evaluation method tested on publicly available dataset
Author Affiliations & Notes
  • Robert Arnar Karlsson
    University of Iceland, Eindhoven, Netherlands
    Oxymap ehf., Iceland
  • Benedikt Atli Jónsson
    University of Iceland, Eindhoven, Netherlands
  • Sveinn Hakon Hardarson
    University of Iceland, Eindhoven, Netherlands
  • Gisli Hreinn Halldorsson
    Oxymap ehf., Iceland
  • Olof Birna Olafsdottir
    University of Iceland, Eindhoven, Netherlands
  • Einar Stefánsson
    University of Iceland, Eindhoven, Netherlands
  • Footnotes
    Commercial Relationships   Robert Karlsson, Oxymap ehf. (I), Oxymap ehf. (P); Benedikt Jónsson, None; Sveinn Hardarson, Oxymap ehf. (I), Oxymap ehf. (P); Gisli Halldorsson, Oxymap ehf. (I), Oxymap ehf. (P); Olof Olafsdottir, None; Einar Stefánsson, Oxymap ehf. (I), Oxymap ehf. (P)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2020, Vol.61, 1638. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Robert Arnar Karlsson, Benedikt Atli Jónsson, Sveinn Hakon Hardarson, Gisli Hreinn Halldorsson, Olof Birna Olafsdottir, Einar Stefánsson; Image quality evaluation method tested on publicly available dataset. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1638.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : The quality of fundus images is an important factor in the successful outcome of diagnosis of various diseases.
We present an updated method for automatically assigning a quality grade to an image. The quality was evaluated using a random forest regression model whose input features had been discovered automatically.
The method was previously tested on two proprietary retinal image datasets were the system outperformed the average human expert. In order to compare the method with other work the method was tested on DRIMDB, a publicly available dataset.

Methods : The method separately evaluates the focus and contrast quality components of a retinal image.
To estimate focus the retinal images are transformed into the Fourier frequency domain.
Contrast evaluation method uses features which were automatically discovered by arranging basic image operators using simulated annealing.
The dataset used for training was provided by Landspítali University Hospital and contained 808 retinal oximetry images and 256 RGB images from a Zeiss FF450 fundus camera. The method was tested on the DRIMDB dataset which consists of 194 retinal images captured with a Canon CF-60UVi fundus camera.

Results : The DRIMDB dataset classifies images into two categories, good and bad. The proposed method was tested on the DRIMDB dataset in two ways. First, the system was trained on all available images from the proprietary datasets. The method was then used to rate the images of the DRIMB dataset. The outcome is shown as a histogram in Figure 1. Secondly, a slight modification to the method was made. A random forest classifier was trained on the images using the contrast and focus features. This setup was trained and tested using 10-fold cross validation. The results are shown in Table I.

Conclusions : The proposed method for evaluation of image quality is more accurate, faster and less subjective than analysis by a single human grader. While differences in testing methodologies make exact comparison with other methods difficult the method performs in line with the best proposed image quality assessment methods.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1. A histogram showing the quality grade for the images in the DRIMDB dataset. The dashed line shows the classification boundary.

Figure 1. A histogram showing the quality grade for the images in the DRIMDB dataset. The dashed line shows the classification boundary.

 

Table I. The proposed method compared with other methods when classifying the DRIMB dataset into good and bad images.

Table I. The proposed method compared with other methods when classifying the DRIMB dataset into good and bad images.

×
×

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.

×