Automating Analysis of the Aorta from Cardiac MR Images

Lead Research Organisation: University of Manchester
Department Name: Computer Science

Abstract

Background
Diseases of the thoracic aorta are common. As a result, tens of thousands of diagnostic and surveillance thoracic aortic Magnetic Resonance (MR) Images are performed annually in the UK.

Unmet needs
1. Aortic measurements made on the scans are time consuming (>20 minutes per scan) and observer dependent, resulting in high costs and inefficiencies.
2. Risk stratification remains poor. Treatment (i.e. surgery) is currently guided by 1-dimensional aortic measurements, which are widely recognised to be inadequate. Parameters that provide more discriminatory and personalised risk stratification are required.

Objectives
To develop novel automated systems which can help clinicians to make diagnostic judgements by providing measurements from the MR images automatically.

In particular the goals of the project are:
1) To develop a statistical shape model of the aorta from annotated MRI images.
2) To develop a system to locate the outline of the aorta in MRI images using the shape model.
3) To evaluate the accuracy of the result, including how well measurements made with the automatic system compare with manual measurements.
4) To analyse the prognostic value of shape parameters and other clinically relevant landmarks from the model to clinical outcomes.
5) To develop multivariate prognostic models combining the shape model parameters with available clinical data using survival analysis methodology.

Proposed project
We have an on-going prospective cohort study of consecutive consenting patients undergoing cardiovascular MRI scanning in Manchester, which includes 6,800 patients as of May 2018 (recruitment 2000/year since Jan 2015). Clinical data, and 'expert-analysed' scan measurements are recorded, together with follow-up data (outcome data from NHS digital, hospital and GP records, follow-up scans).

1) Methodology development
The student will develop a system to segment the aortic root and thoracic aorta from MRI images. The system will be trained using data from approximately 1000 patients (including normal and abnormal aortas). The system will build on state-of-the-art shape modelling and matching techniques developed in Manchester [1] augmented with Convolutional Neural Networks where they can be shown to be more robust. By locating clinical landmarks and the outline of structures of interest we will be able to automate the standard 1D clinical measurements and to produce novel new parameters describing the shape and appearance in more detail.

2) Validation
The analysis models developed in Part 1 will be validated on data from a second group of 1000 patients, including patients normal and abnormal aortas. Accuracy of automatic measurements will be compared with those produced by human experts.

3) Prognostic utility and modelling
The utility of scan analysis parameters as prognostic factors (for a combined outcome of increase in aortic size, aortic intervention and death) will be evaluated using survival analysis methodology. Multivariable prognostic models, incorporating clinical variables (such as diagnosis [e.g. syndrome connective tissue disorders and non-syndrome disorders] and other risk factors [e.g. hypertension]) and scan analysis parameters will be developed.

EPSRC Research Areas: Image and Vision Computing, Heathcare Technologies

Publications

10 25 50