Abstract
Purpose :
Optical coherence tomography (OCT) in the remote care setting can be challenging due to self-scanning limitations. Patient-operated OCT systems need to detect the state of patient alignment and determine the OCT image quality for fully automated imaging. Here, we demonstrate an algorithm that automatically assesses the OCT image quality using machine learning.
Methods :
A total of 141 volumes were acquired with a prototype low-cost OCT system from 18 subjects using either one or both eyes. The size of each volume was 7.00 mm x 5.78 mm x 2.77 mm and all volumes were centered on the fovea. For each volume, multiple feature maps were created based on signal strength, signal to background ratio, and individual A-scan contrast. These feature maps were then combined into a joint probability map previously presented (Elezaby et al., IOVS, 2020). By placing a threshold on the joint probability map, we generated groups of likelihood functions as inference models for each of the feature maps. The OCT quality map was formed using the likelihood functions and a posterior probability from the inference model. The OCT scan quality was assessed using the ETDRS grid consisting of three concentric circles (radii of 0.5 mm, 1.5 mm, and 2.9 mm) divided into nine subfields. An OCT volume was considered poor quality based on two criteria. The first involved three or more central and inner ETDRS subfields having low confidence. The second involved two or more ETDRS subfields that include at least one outer subfield having low confidence. OCT volumes were assessed manually by an experienced grader to obtain the ground truth.
Results :
Out of the 141 volumes, there were 90 true positives (good quality), 44 true negatives (poor quality), 4 false positives, and 3 false negatives. Fig. 1 shows an example for each scenario. The algorithm yielded a sensitivity of 0.97 with 95% confidence interval (CI) [0.91, 0.99], specificity of 0.92 with 95% CI [0.80, 0.98], and success rate of 0.95 with 95% CI [0.90, 0.97].
Conclusions :
We demonstrated a new machine learning method that automatically grades OCT volumes with high sensitivity and specificity using the ETDRS grid. This algorithm could help acquire good quality data in the remote care setting.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.