Abstract
Purpose :
Accuracy of automated retinal layer segmentation is dependent on image quality. The purpose of this study is to assess the feasibility of an automated quality SD-OCT assessment tool through a “segmentation confidence score” as a predictor of image quality and segmentation accuracy.
Methods :
Eighty-six SD-OCT scans (Cirrus, Zeiss) were included in this study. Following review of each of the 128 slices, two expert readers assigned each scan a quality grade based on manual segmentation feasibility: good - gradable with minimal defects, fair - gradable despite notable defects, or poor - ungradable. Using an automated image feature assessment platform, traditional image quality parameters related to signal intensity including signal intensity mean, median, variance, skew, kurtosis, and homogeneity were calculated for each slice and averaged for each scan.
Utilizing an automated machine learning (ML) enhanced retinal layer segmentation platform, the following layers were segmented: internal limiting membrane, outer nuclear layer, ellipsoid zone, retinal pigment epithelium, and Bruch’s membrane. A segmentation confidence score for each layer of each slice was obtained, ranging from 0-100 based on the percentage of x-axis pixels where the ML platform was able to detect a layer. Confidence scores for all layers and slices were averaged to obtain a mean confidence score for each scan. The performance of a random forest ML classifier was assessed with traditional image quality parameters and also with the segmentation confidence scores. AUC values were obtained from 10 iterations of 5-fold cross-validation [reported: mean (variance)].
Results :
Using the image parameters related to signal intensity measures alone, the AUCs for classifying good, fair, and poor quality scans were 0.951 (0.001), 0.874 (0.005), and 0.931 (0.007), respectively. The addition of the ML-derived segmentation confidence scores to the ML scan quality classifier increased the AUCs for good, fair, and poor quality scans to 0.975 (0.0005), 0.929 (0.003), and 0.992 (0.0001), respectively.
Conclusions :
The ML-based segmentation confidence score enabled a highly discriminating tool for assessing OCT quality. Automated quality assessment could enable rapid feedback for optimizing automated image analysis, clinical trial inclusion, and clinical feedback for photographers.
This is a 2021 ARVO Annual Meeting abstract.