Applications are invited for a fully funded 3,5 year full-time PhD studentship (including home tuition fees, annual stipend and consumables) starting on 1st February 2024.
Aim of the project
Cardiovascular disease is the most common cause of death worldwide. Detection of heart damage early is important to identify which patients are at risk. Imaging such as with Cardiac MRI is potentially well suited to detect damage at an early stage, and may allow preventative therapy to be started.
‘Strain’ is a potentially important biomarker of early damage. Strain is the amount of shortening and lengthening of heart muscle in each region of the heart. This project will develop novel artificial intelligence (AI) methods to automatically extract strain from standard MRI clinical imaging scans. The aims are: 1) develop tools for robust quantification of local strain; 2) train AI methods to predict accurate strain from standard MRI exams, and 3) use strain to predict outcomes in patients with heart disease.
We anticipate that this project will lead to implementation of AI driven solutions to enhance patient diagnosis and treatment. In particular, this project will enable better management of patients undergoing chemotherapy for cancer, who need constant monitoring to see if heart function is impaired due to cancer therapy and for detection of early-stage heart failure.
This is a collaborative project between the Biomedical Engineering and Imaging Sciences department of King’s College London and the heart failure and cardio-oncology unit at Royal Brompton Hospital.
Project description
Myocardial strain is a fundamental clinical biomarker which directly quantifies local tissue performance in each region of the heart. Currently, the most accurate method of measuring strain is to acquire specialised images, which adds time to already lengthy imaging scan. A simpler method is to predict strain from standard images which are already acquired in all imaging exams, as part of standard routine clinical care. This would save costs and reduce scan times and allow widespread clinical application. However, current methods for estimating strain from standard images do not work well for regional assessments. This project will utilize novel AI methods for estimating strain from standard images, by learning the relationships between image motion features and strain.
METHODOLOGY
Year 1: The first year will include a literature review on strain quantification from images. The student will also attend several clinical MRI scan sessions and learn how the images are acquired. An initial paper will compare results using standard methods in 20 healthy volunteers, as part of a reproducibility study performed at St Thomas’ Hospital. State of the art machine learning and AI algorithms will be then developed for estimation of strain from “MRI tagging” images which are specifically designed to image accurate motion patterns (using a special type of imaging which is sensitive to displacements). The methods for strain estimation from tagging images will improve upon previous work done using UK Biobank tagging data in 5,000 participants of the UK biobank [1], for whom manual ground truth is available. These methods will be extended using physics based motion analysis, thereby imposing physical constraints on the derived displacements (including incompressibility of the 3D displacement field). Repeatability will be assessed using scan-rescan comparisons in tagged and untagged images.
Year 2: Using existing UK Biobank study approvals, transfer learning AI methods will be developed to take advantage of a large amount of MRI information available in ~60,000 participants. These scans will be used in a supervised multimodal data analysis to predict strain from standard MR images. The results of the MR tagging analysis will be used as ground truth to train a network to derive strain from standard cine images. Physics based constraints will be imposed to make sure the resulting strain is physiologically reasonable. A benchmark comparison will be derived from previously published methods including [2].
Years 3 and 4: The methods will be applied to patients previously imaged at St Thomas’ Hospital with heart failure from coronary artery disease. In 700 patients with heart disease, standard cine MRI images were acquired and outcomes including death or hospitalization for heart failure over 2-4 years were recorded [3]. The ability of accurate regional strain measures to predict adverse outcomes will be tested by examining the additional predictive performance of strain to standard measures of mass volume and muscle viability and existing methods of strain quantification.
[1] Ferdian E et al Radiology: Cardiothoracic Imaging 2(1):e190032 (2020)
[2] Loecher M et al Medical Image Analysis 74:102223 (2021)
[3] Perera D et al New England Journal of Medicine 10.1056/NEJMoa2206606 (2022)
Informal email enquiries from interested students to the supervisors are encouraged (contact details below).
Dr Sohaib Nazir - [Email Address Removed]
Prof. Alistair Young - [Email Address Removed]
Applications
Please visit the studentship website for more information on eligibility criteria and application instructions.
Continue with Facebook
