Myocardial infarction (MI) (a.k.a. heart attack) occurs when the supply of blood to the heart muscle is suddenly blocked. In 2013/14 over 80,000 patients were admitted to hospital with an acute MI in England and Wales.
Cardiovascular Magnetic Resonance (CMR) can detect and quantify myocardial infarction with unique precision using the late gadolinium enhancement (LGE) technique. This can provide prognostic information independently of ejection fraction. However, CMR is currently dependent on the limited availability of qualified experts, subjective assessment and is too time-consuming for clinical practice. Semi-automated methods (such as Full-Width Half-Maximum, N-standard deviations, etc.) also rely on the subjective determination of endocardial/epicardial borders. On top of this, the above classifiers also fail to generalize to datasets acquired by different labs due to acquisition-related variation (imaging protocols, field strength, vendor etc.). The presence of artefacts is another challenge for this type of algorithms.
Deep learning (DL) is a rapidly growing trend in general data analysis that currently drives the artificial intelligence (AI) boom. Deep convolutional neural networks (CNNs) are DL architectures and related algorithms which are well-suited for image analysis tasks.
The project will use DL to develop tools to enable fast, accurate, automatic 3D model generation of myocardial scar data. Data from 1,958 research patients who have had CMR following MI will be used. The datasets for this project were obtained at 7 leading CMR centres in UK. The tools developed in this project will allow multi-centre comparisons, which in turn, will facilitate the establishment of novel CMR-based biomarkers (infarct size) as predictors of hard clinical points. The proposed algorithms will also allow high-risk sub-groups to be identified for future interventional studies.
The research will be driven by the following hypotheses:
• CNNs will replicate a cardiac MRI expert manual labelling of myocardial scar tissue in cardiac MRI datasets of MI patients, but be faster and more robust.
• AI-based techniques of scar quantification will be more strongly associated with clinical outcomes, compared with the standardised semi-automated techniques.
• Deep learning can be used to learn CMR acquisition-invariant representations, so that a (scar tissue) classifier, that has been trained on data from a specific scanner vendor/field strength/protocol, can be applied to data that was differently acquired.
• Deep learning architectures can be trained to recognise artefacts in the remote myocardium.
Data Science at School of Science and Technology, NTU is equipped with powerful hardware for achieving incredible performance in deep learning problems.
Applicants must apply using the online form on the University Alliance website at https://unialliance.ac.uk/dta/cofund/how-to-apply/
. Full details of the programme, eligibility details and a list of available research projects can be seen at https://unialliance.ac.uk/dta/cofund/
The final deadline for application is 12 April 2019.