Purchase this article with an account.
Carla Agurto Rios, Honggang Yu, Jeff Wigdahl, Victor Murray, E Simon Barriga, Wendall Bauman, Peter Soliz; DETECTION OF NEOVASCULARIZATION IN DIGITAL FUNDUS IMAGES USING A MULTISCALE ANALYSIS APPROACH. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5520. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To present a software system that automatically detects the presence of neovascularization in the optic disc (NVD) and elsewhere in the retina (NVE). This system analyzes retinal vessel properties such as tortuosity, width, local density, and image representations at different spatial frequency scales in order to determine if the vasculature is abnormal due to the presence of neovascularization.
N=350 digital fundus photographs were collected retrospectively from the University of Texas Health Sciences Center in San Antonio and the Retina Institute of South Texas. The distribution of the images is only NVD (N=50), only NVE (N=50), both NVD and NVE (N=50), and normal cases (N=200). The algorithm for detecting neovascularization is based on adaptive vessel segmentation and a multiscale analysis approach. To detect neovascularization, the images were divided into two regions of interest (ROI). The first ROI is a circle of 1 disc diameter centered in the optic disc, which is used for detecting NVD. The remaining part of the retina, which constitutes the second ROI, is used for detecting NVE. We applied an adaptive vessel segmentation methodology to segment the new vessels in order to avoid over-segmentation. Features for analysis included fractal dimension index, size distribution with granulometry, and frequency components at different scales obtained using amplitude-modulation frequency-modulation. These features were the input to a support vector machine (SVM) classifier that separated the images into three classes: normal, NVD, and NVE.
We tested our system using 10-fold cross-validation. An area under the ROC curve (AUC) of 0.93 and sensitivity/specificity of 0.95/0.68 were obtained for the classification of NVD. For the detection of NVE, the system achieved an AUC of 0.91 and sensitivity/specificity of 0.92/0.71.
This work presents a novel methodology for the automatic detection of NVD and NVE in digital fundus images. Our method achieved the objective of characterizing the vasculature into normal, NVD, or NVE classes. The proposed adaptive vessel segmentation method adequately detects the minute vessels present in neovascularization, allowing us to correctly extract features of these structures, thus minimizing the inclusion of spurious information from other structures in the retina.
This PDF is available to Subscribers Only