Machine Learning in Cardiovascular Imaging

For patients with heart disease, special medical images called cardiac magnetic resonance (cMRI) are used to diagnose and plan treatments. Within one session with cMRI many images are taken, which can provide a lot of information on the function of the heart, the heart muscle characteristics and the volume of the heart chambers. Currently, only a small part of all the information in the images is used by clinicians/doctors. 

For the information that is used, clinicians currently manually draw circles on selected images to obtain measurements.  Using this manual technique, a clinician would need to annotate over 1000 images from one cMRI scan to gather all relevant or helpful information. This is un-realistic, which often means only key analysis takes place.  Automating the collection and analysis of all scan info both removes the laborious task of manually annotating thousands of images while also improves the quality of the information gathered, by processing information on the scan that is not easily accessible by clinicians. 

This project uses AI to assist doctors in assessing heart function from medical images, specifically cardiac magnetic resonance (cMRI).  These special medical images are useful in the diagnosis and treatment plans for patients with heart diseases. By increasing the information gained from cMRI images through automation, guidance on treatment will be improved and better care can be provided to patients. Removing repetitive and laborious tasks will also provide doctors with more time to assess complex cases and other very important parts of patient care. By extension, this automation will reduce the number of hospital admissions for heart failure, which is currently one of the largest NHS hospital costs.  

This projects automated method for the analysis of cMRI scans will build on an existing set of AI tools (developed for the UK Biobank population study) that act together to analyse images and provide new, very accurate measures of cardiac function. This existing method could be improved to ensure that it works on a large variety of patients seen in clinical cardiology departments such as those at Guy’s and St Thomas’ Trust (GSTT). Therefore, the group of AI tools will be tested on a large amount of data that represents a varied patient population.  

As part of testing on this larger population, we will evaluate which measurements (or set measurements) best predicts mortality and responses to treatment in subgroups of patients. This will also be compared to other forms of cardiac imaging that are often used to make sure these tools are better than the existing methods.   

The output of this project will enable tailored information about prognosis and treatments in the future. By improving guidance on treatment, patients with cardiac health issues will be better cared for and doctors will be freed up to provide better care for patients.