Leveraging Modelling and Machine Learning for the Analysis of Curvilinear Structures in Medical Images

  • Roberto Annunziata

Student thesis: Doctoral ThesisDoctor of Philosophy


Recent clinical research has highlighted important links between a number of diseases and the tortuosity of curvilinear anatomical structures such as corneal nerve fibres, suggesting that tortuosity changes might detect early stages of specific conditions. Currently, clinical studies are mainly based on subjective, visual assessment, with limited repeatability and inter-observer agreement.

In this thesis I have endeavoured to address these problems by proposing a fully automated framework for image-level tortuosity estimation, consisting of a hybrid segmentation method and a versatile tortuosity estimation algorithm. The former combines an appearance model, based on a Scale and Curvature Invariant Ridge Detector (SCIRD), with a context model, including multi-range learned context filters. The latter is based on a novel tortuosity estimation paradigm in which discriminative, multi-scale features can be automatically learned for specific anatomical objects and diseases.

I have validated each module of the system separately and then assessed their impact on the tortuosity estimation performance (target application). The segmentation module has been tested on 5 challenging data sets, including corneal nerve fibres (not public, provided by our clinical collaborators at MEEI, Harvard Medical School, USA), neurites (2 benchmark data sets) and retinal blood vessels (2 benchmark data sets). The tortuosity estimation module has been validated on a data set including 140 corneal nerve images, the largest ever used for this task, to my best knowledge.

Experimental results show that (1) the segmentation module outperforms state-of-the-art hand-crafted and hybrid approaches; (2) the tortuosity estimation module performs better than state-of-the-art and widely used tortuosity indices; (3) the whole system matches and sometimes even exceeds tortuosity estimation performance of experienced observers when compared against each other, a level of performance that will allow us to deploy the system on much larger data sets, with the aim of discovering new links between tortuosity and specific diseases in an objective and repeatable fashion.
Date of Award2016
Original languageEnglish
SupervisorManuel Trucco (Supervisor)

Cite this