Purchase this article with an account.
Minhaj Nur Alam, Damber Thapa, Jennifer I Lim, Dingcai Cao, Xincheng Yao; Automatic classification of sickle cell retinopathy using quantitative features in optical coherence tomography angiography. Invest. Ophthalmol. Vis. Sci. 2017;58(8):1679.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Automatic detection and quantitative classification are desirable for effective diagnosis of sickle cell retinopathy (SCR). This study is to explore automatic detection and classification of SCR by characterizing features in optical coherence tomography angiography (OCTA) images.
OCTA images of 35 sickle cell disease (SCD) patients (12 males and 23 females; 35 African Americans) and 14 control subjects (11 males, 3 female, 5 African Americans) were used. The mean age was 40 years (range 24 to 64) for the patients and 37 years (range 25 to 71) for the controls. OCTA images of both eyes were analyzed, so the database consisted of 70 SCD and 28 control eyes. Seven feature vectors, including blood vessel density, vascular tortuosity, diameter, vessel perimeter index, foveal avascular zone (FAZ) area, contour irregularity of FAZ, and parafoveal avascular density were calculated from the OCTA images. Three classifiers, i.e., support vector machine, k-nearest neighbor algorithm and discriminant analysis, were used to classify the OCTA images. For SCR vs. control classification, the algorithms used a random 50% of OCTA images as a training set and the rest (50%) of the images as test set in each simulation. For interstage classification (mild vs. severe) among SCR patients, 95% of the data were used randomly to train the classifier to predict the rest of the 5% data correctly. Sensitivity, specificity, and accuracy were calculated to examine the performance of the algorithms.
For SCR vs. control case, all three classifiers perform well with an average accuracy of 98% using the optimized feature vectors. For inter-stage classification, support vector machine shows better performance compared to the other classifiers. Table 1 shows the performance of each classifier in terms of sensitivity, specificity, and accuracy. Among all 3 classifiers, support vector machine shows the best performance with 100% sensitivity, 100% specificity and 100% accuracy for SCR vs. control classification and 97% sensitivity, 98% specificity and 97% accuracy for inter-stage classification.
The automated classification algorithm with quantitative feature vectors can successfully predict SCR and identify the stage by analyzing OCTA images. This shows the effectiveness of the feature vectors calculated from OCTA images for automatic classification of SCR.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only