Purchase this article with an account.
I. Benche, S. Madden, S. Nemeth, S. Russell, M. Verdugo–Gadzik, S. Wolf, P. Soliz; Assessment of Fill–Rate Statistics to Automatically Segment and Classify Choroidal Neovascular Lesions in Age–Related Macular Degeneration . Invest. Ophthalmol. Vis. Sci. 2005;46(13):1564.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: The fill–rate characteristics as interpreted from fluorescein angiography (FA) videos are used clinically as the basis for the classification and treatment of choroidal neovascularization (CNV) associated with age–related macular degeneration (ARMD). The objective of this project was to assess the utility of fill–rate characteristics for automatically classifying the sub–types of CNV and determining its margins. Methods: Utilizing Retinaview, a Java based software package, with pre–registered, digital FA videos (512 X 512 pixels, 50–110 frames), an expert retinal grader (MD) annotated the boundaries of CNV lesions from a natural history study of a longitudinal cohort of patients with occult CNV from ARMD. Forty FA videos from 25 subjects labeled with occult FPED (fibrovascular pigmented epithelial defect) subtype of CNV were analyzed. Parameterized images, such as the initial and peak fluorescence intensity, the time to onset of fluorescence, and the rate of filling, were obtained from the FA videos with two goals in mind: to automatically locate the margins of the CNV and to determine the subtype of the CNV lesion (classic or occult). The images were divided using a 7x7 grid centered on the macula, on which textural features (such as contrast, periodicity, texture energy measures) are computed. A linear discriminant classifier was trained using as ground truth the percentage of lesion in each of the grid boxes, and a cross–validation technique was used to asses its performance. Results: An average of 43 out of 49 boxes (88%) were correctly classified by the percentage of lesion from each of the boxes (2 classes: 0–50%, 51–100% lesion). A second classifier was developed to separate the lesions into occult or classic CNV subtypes for the 11 cases out of 40 where the CNV lesions also contained the classic subtype. A specificity of 0.62 and a sensitivity of 0.67 were achieved when defining the number of pixels correctly classified as classic lesion as true. Conclusions: This project has demonstrated that the parameters obtained by extracting the fill rate characteristics from FA videos show great potential in the development of an automatic CNV segmentation algorithm. A larger dataset and more information given to the algorithm (adding FA still images) should result in improved performance of this methodology.
This PDF is available to Subscribers Only