Retinal Vascular Segmentation Using Superpixel-based Line Operator and Its Application to Vascular Topology Estimation.
Tong Na,Jianyang Xie,Yitian Zhao,Yifan Zhao,Yue Liu,Yongtian Wang,Jiang Liu
DOI: https://doi.org/10.1002/mp.12953
IF: 4.506
2018-01-01
Medical Physics
Abstract:PurposeAutomatic methods of analyzing of retinal vascular networks, such as retinal blood vessel detection, vascular network topology estimation, and arteries/veins classification are of great assistance to the ophthalmologist in terms of diagnosis and treatment of a wide spectrum of diseases. MethodsWe propose a new framework for precisely segmenting retinal vasculatures, constructing retinal vascular network topology, and separating the arteries and veins. A nonlocal total variation inspired Retinex model is employed to remove the image intensity inhomogeneities and relatively poor contrast. For better generalizability and segmentation performance, a superpixel-based line operator is proposed as to distinguish between lines and the edges, thus allowing more tolerance in the position of the respective contours. The concept of dominant sets clustering is adopted to estimate retinal vessel topology and classify the vessel network into arteries and veins. ResultsThe proposed segmentation method yields competitive results on three public data sets (STARE, DRIVE, and IOSTAR), and it has superior performance when compared with unsupervised segmentation methods, with accuracy of 0.954, 0.957, and 0.964, respectively. The topology estimation approach has been applied to five public databases (DRIVE,STARE, INSPIRE, IOSTAR, and VICAVR) and achieved high accuracy of 0.830, 0.910, 0.915, 0.928, and 0.889, respectively. The accuracies of arteries/veins classification based on the estimated vascular topology on three public databases (INSPIRE, DRIVE and VICAVR) are 0.90.9, 0.910, and 0.907, respectively. ConclusionsThe experimental results show that the proposed framework has effectively addressed crossover problem, a bottleneck issue in segmentation and vascular topology reconstruction. The vascular topology information significantly improves the accuracy on arteries/veins classification.