Abstract
Purpose :
Interpretation of optical coherence tomography (OCT) and OCT angiography (OCTA) images is largely subjective, and screening all diabetics for diabetic retinopathy (DR) is a major public health challenge. We sought to apply machine learning techniques to standardize and automate the diagnosis of nonproliferative diabetic retinopathy (NPDR) using OCT and OCTA.
Methods :
This was a two-center, retrospective, cross-sectional imaging study. Inclusion criteria included age > 18 and a diagnosis of diabetes mellitus (DM). Exclusion criteria were a diagnosis of a retinal disease other than DR, media opacity precluding imaging, and high myopia. Patients underwent a full dilated eye exam, and OCT and OCTA imaging with the Zeiss Cirrus Angioplex. Machine learning techniques were then applied. Three pathophysiologically important features were extracted from each layer of the OCT: reflectivity, curvature, and thickness. Four features were extracted from the OCTA: blood vessel caliber, vessel density, the size of the FAZ, and the number of bifurcation and crossover points. Data from these seven extracted features were then fed into a random forest classifier to train and test the system via two-fold and four-fold cross validation. These results were then compared to the clinical grading of NPDR, which was considered the gold standard.
Results :
82 patients with DM II were included in the study (age range 18-79, 54% female). By clinical exam, 26 patients had no DR, and 56 had mild, moderate, or severe NPDR. When OCT images alone were analyzed by the system, accuracy of diagnosis was 92%, sensitivity 84%, and specificity 100%. When OCTA images alone were analyzed, accuracy was 97%, sensitivity 98%, and specificity 96%. When data from the two imaging modalities were combined, the system achieved 100% accuracy, sensitivity, and specificity. Subgroup analysis of different OCT features further revealed that the curvature of the inner nuclear layer, the reflectivity of the myoid zone, and the thickness of the nerve fiber layer had the greatest value in discriminating between normal and NPDR cases.
Conclusions :
Automated diagnosis of NPDR using a combination of OCT and OCTA images is feasible and highly accurate. As these technologies become less expensive and more ubiquitous, this may allow for broader screening of diabetic patients, particularly in areas lacking a sufficient workforce.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.