Automated MRI Imaging

Measuring cardiac structure and function by cardiac imaging drives most clinical decisions in cardiovascular medicine.

Cardiac function is best measured using cardiac magnetic resonance (CMR) imaging, but even a basic examination can take up 60 minutes for a basic examination making it an expensive investigation. Once the images have been captured, they need expert analysis by a clinical expert. This expertise is rare, expensive, and most importantly, variable. The variability can lead to errors in clinical-decision making, which can have profound effects on patients.

The team aimed to tackle the problem by using AI to:

• Speed up the scan time to lead to reduced waiting lists and costs, more availability at more centres, and a more pleasant exam for patients.
• Automate the analysis of the scanned images. This would reduce the waiting time for the final report, reduce the need for expert clinician time, reduce costs and improve treatment choice for patients.
• Move towards patient-specific recommendations. By making patient-specific reference ranges, we can categorise/diagnose patients with higher precision.

The project was delivered through 6 key steps:

Precision AI development

The team built on AI that that had been previously developed to automate the analysis of cardiac MRI, to improve measurement precision, reduce the time waiting for the report and freeing up clinicians.

Protocol optimisation

Cardiac MRI protocols are often treated as ‘one size fits all’.  The team set about to tailor the acquisition sequence to get rid of unnecessary images.

Clinical delivery

The AI was implemented directly on the MRI scanner so that analysis was performed as the images were being acquired. This aim was to complete the analysis by the time that the patient had left the scanning room.

Patient engagement

The team formed a focus group to gather patients’/the public perception of AI, their thoughts on using it for image acquisition, and how it should be presented. This improved our understanding of patient perception allowing the team to tailor their approach (e.g. terminology used, patient information leaflets) to patients.

Clinician engagement

The team also engaged with the reporting clinicians to see what worked best for them, and to make sure that the clinical workflow was optimised.

Patient-specific reference ranges

Patient ranges specific to the AI were created by using data from 4500 subjects from the UK biobank and 2000 additional healthy volunteers, as well as sex, age, size, and ethnicity.

The fund focused on innovative approaches to care for heart failure patients and this project addressed that in four key approaches:

• Clinical deployment of AI in Cardiac MRI: The project used AI to automate and enhance the precision of cardiac MRI scans which has not been widely applied to actual patients in the way this project has.
• Benchmarking Against Human Precision: The project established a unique benchmarking process to evaluate the precision of the AI against human precision. This provided a tangible way to measure the success and accuracy of the AI in a clinical setting.
• Multi-Pathology Validation: The AI was validated using a dataset that contained multiple pathologies. This approach ensured that the AI could handle a wide range of cardiac conditions, increasing its potential usefulness in real-world clinical settings.
• Clinical validation: Even as the project leaned heavily into AI automation, it retained a degree of human supervision to validate the scan, slice prescriptions, and reports. This novel hybrid approach ensures patient confidence and maintains a high level of accuracy.

Funding from the BHF Hope for Hearts Fund made a difference in the following way

• Clinical delivery: most grants encourage the development methodologies and technologies. The Hope for Hearts Fund allowed the team to concentrate on clinical delivery, allowing them to get the technology to the patient.
• Holistic approach: as well as clinical delivery, the team were able to evaluate its impact on both patients and reporting clinicians.
• Training clinical AI scientist: The funding enabled the training of a clinical radiologist in developing the AI