Purchase this article with an account.
Esra Ataer-Cansizoglu, Veronica Bolon-Canedo, Deniz Erdogmus, Katherine Abrahams, Susan Ostmo, Robison Vernon Paul Chan, Jayashree Kalpathy-Cramer, Michael F Chiang; A GMM-based Feature Extraction Technique for the Automated Diagnosis of Retinopathy of Prematurity. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):5252.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Plus disease is an important factor for identification of severe retinopathy of prematurity. Arterial tortuosity and venous dilation in the retina are important signs for determining plus disease. In order to build a computer-aided diagnosis system, it is necessary to extract these features from vessel points or segments. Then, an image is represented with "regular" statistics (e.g. minimum, maximum or mean of these values). However, these statistics provide biased estimates as an image contains both healthy and abnormal vessels. Moreover, one should distinguish between arteries and veins to account severity properly. We have developed computer-based image analysis methods to (1) overcome these limitations with a novel feature extraction method that represents each image with the parameters of a two-component Gaussian Mixture Model (GMM), (2) compare the classification accuracy for the features extracted from the whole image versus features extracted from arteries and veins separately.
77 retinal images were manually segmented by an expert, and arteries and veins were manually annotated. Reference standard diagnoses were provided for each image by a consensus of 3 experts, together with the actual clinical diagnosis. Manual segmentations were fed into our image processing system that yielded the proposed GMM statistics and regular statistics of 10 image-based features. We built an automatic diagnosis system for identification of plus disease vs. preplus vs. neither with various classifiers using (i) GMM statistics, (ii) regular statistics, and (iii) both of them on features extracted from the whole image and on features extracted from arteries and veins separately.
The classifiers trained with the proposed features outperformed the classifiers using regular statistical features. The best classification accuracy among different classifiers was 90.5% with the proposed features and 88.2% with the regular statistical features. Extracting features from the whole image without distinguishing veins and arteries gave better accuracy (90.5%) compared to using features from arteries and veins separately (80.2%).
Overall tortuosity of an image can be well represented with a two-component GMM of the image-based features. GMM statistics are useful for building accurate automated diagnosis systems and they remove the necessity of distinguishing arteries from veins.
This PDF is available to Subscribers Only