Purpose:
The quality of optical coherence tomography (OCT) images has a significant impact on measurements and diagnosis. The device provides an overall score for signal strength, but when only part of the image is poor, global quality metrics might not suffice. The purpose of this study was to develop an automated method for determining the quality of individual A-scans, in a way that is insensitive to pathology.
Methods:
Three OCT experts independently labeled the quality of each A-scan in 270 images. There were 90 images each of the macula, optic nerve head, and circumpapillary retinal nerve fiber layer. To test robustness to pathology, one-third of the eyes had no glaucoma, one-third had early glaucoma, and one-third had advanced glaucoma. Using the experts’ labels, we trained a hierarchy of support vector machines (SVM) at multiple scales and used histogram-based metrics. Specifically for the SVM, we projected different sized groups of A-scans into a high dimensional space, then compute the hyperplane that minimizes classification error. To improve the classification performance, the scans were normalized to be insensitive to various effects, such as pathology and eye movement during acquisition.
Results:
There was strong agreement between the experts and our new algorithm. The following chart is the percent agreement between each of the three experts, their mode, and our new algorithmFor our algorithm, the area under the receiver operating characteristic curve (AUC) was 0.94 for the macula, 0.96 for the optic nerve head, and 0.95 for the retinal nerve fiber layer.
Conclusions:
We show that our automated algorithm performs on par with experts in discriminating local OCT image quality. This localized assessment can be used to improve the performance of segmentation and image fusion algorithms.
Clinical Trial:
www.clinicaltrials.gov NCT00286637
Keywords: imaging/image analysis: clinical • image processing