Abstract
Purpose :
For screening purposes, fundus imagers are often placed in settings with inexperienced operators. Although robotic fundus imagers can improve the success rate for such operators, having immediate feedback on the quality of acquired images can allow even an inexperienced operator to retake an image if needed. The purpose of this study is to evaluate an image quality algorithm on a subset of data taken as part of a clinical screening program.
Methods :
This was a retrospective review of patient fundus images performed with the NW400 (Topcon, Tokyo, Japan) in 580 primary care offices across the United States. Images were obtained in primary care offices and sent via the cloud for interpretation. Images were read/interpreted by licensed eye care professionals in the Topcon Screen Reading Center via the Topcon Harmony RS PACS system. A subset of images was extracted and processed with a convolutional neural network (CNN) trained on NW400 images (Topcon Corporation, Tokyo, Japan) which had been broken into 9 equal segments. The CNN reported on the image quality of each segment with a score that ranged from 0 to 1. For each image, the results of the nine segments were averaged to obtain the overall score. The cutoff for poor quality was 0.995.
Results :
A subset of 229 images from 229 eyes of 229 subjects was evaluated. Figure 1 shows an example of the output from the CNN on a fundus image. Only 22 (9.6%) of the images were Fair or Unreadable, with the rest Good or Excellent. The accuracy of the CNN was 0.97, with a sensitivity of 0.90 and a specificity of 0.97. The average run-time of the algorithm was better than 1 second on an NVIDIA Jetson TX2, although the load time was 5 seconds. Run time on Intel® Core™ i7-8700 CPU @ 3.20GHz 3.19 GHz was within 2 seconds.
Conclusions :
A CNN was able to provide a rapid assessment of image quality that agreed well with the assessment of professional graders in data from a real-world screening program.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.