Multiscale tubular features and representation

Font Calvarons, Adrià

Multiscale tubular features and representation

License

Citation

Font Calvarons, A. Multiscale tubular features and representation. 2017. handle: http://hdl.handle.net/10230/28343

This citation was generated automatically.

Abstract

This project aims at building an automatic vessel segmentation pipeline to be used in medical images. Such a goal is shared by many disciplines within medicine, facing considerable amounts of images containing tubular structures. This thesis rst reviewes and discusses well-known multiscale approaches in vessel enhancement, intended for vessel segmentation. In particular, the Frangi, Optimally Oriented Flux and Steerable lters are discussed. They rely on the assumption of local tubular geometry of vessels. However, this assumption can often result in a poor description of vessels in medical images since these usually contain branchings, crossings and irregular vessel structures. For this reason, this work further explores the feasibility of combining these lters' outputs into a feature set, used in a Support Vector Machines scheme in order to make a binary classi cation discriminating vessel regions from the background. An additional preprocessing step is also added in order to prompt tubularity measure performance and normalize the intensity range of di erent input images. The pipeline is implemented using ITK and OpenCV, and has been tested on DSA images from di erent anatomical locations. Results are qualitatively and quantitatively compared to segmentation produced by several thresholding methods on preprocessed images, in order to assess its accuracy. The current pipeline is exible to explore additional application speci c features. This forms part of the future work of this project, which also includes more thorough evaluation of the pipeline.