Medical Devices

An AI doctor is analysing heart scans in dozens of hospitals

An MRI scan of a heart
Chris Baraniuk

In a dimly lit room full of computers at St Bartholomew’s Hospital in London, doctors pore over images of people’s hearts. Until recently, medical staff here had to interpret the blotchy on-screen images purely by sight. Now artificial intelligence is helping to explain what they are looking at.

Charlotte Manisty, a consultant cardiologist at St Bartholomew’s, analyses an MRI scan of a struggling heart and points to blue smudges over one area of muscle. The image on her screen has been coloured in by AI. A swathe of blue around the left ventricle, the heart’s main pumping chamber, means that not enough blood is getting to that part of the muscle.

The volume of blood reaching each bit of the heart is a good indicator of how well it is functioning. Not only does the AI give a colour-based indicator, it also provides a numerical estimate of blood flow for each region too. Previously, doctors had to eyeball black and grey scans to make a judgement about how much blood was present. Getting an actual number needed specialists and took several hours or days.

“All of the things that we’re working on here are to try and reduce the training required,” says Manisty. The AI works completely automatically and delivers its analysis in around two minutes, she says.

The same system is now used at more than 30 hospitals worldwide and has analysed more than 20,000 MRI scans to date. It was developed by Peter Kellman and Hui Xue at the National Institutes of Health in Maryland and their colleagues.

To get the algorithm to correctly identify each bit of the heart in MRI scans, the team trained it on more than 1900 scans of around 1000 patients. The system was then tested against 200 scans from 105 patients to show that it could reliably select each area of heart muscle. It proved to be at least 90 per cent accurate in each case. The system was also previously trained to quantify blood flow and compared against cardiac positron emission tomography, where it was found to be 92 per cent in agreement with that method.

The team plans to upgrade the AI soon so that it can actually determine a patient’s condition, for instance by stating whether it thinks it has spotted a blocked artery, diseased tissue or a healthy heart.

Manisty says the tool has been “a massive step forward” that has given her team greater confidence when reading scans. Along with other algorithms that Kellman and Xue have developed to, for example, allow scans to continue even when patients aren’t easily able to hold their breath, AI has improved efficiency at St Bartholomew’s, she says. Previously, the department scanned around 25 people a day – now it is well into the 30s.

Shehab Anwer at University Heart Center Zurich questions whether the colour-coding could obfuscate certain features of a heart scan, perhaps meaning that doctors miss other signs of disease. Manisty says that the original, grey scans are all still accessible in the system.

William Bradlow at Queen Elizabeth Hospital Birmingham says that there is little risk of distorting the scans. Interpreting MRI images of hearts is tricky, he adds, but with help from AI, more doctors could be doing it on a regular basis.

“This will make assessment of what has previously been quite a difficult and subjective activity that much easier,” he says.


Related posts

AI improves chest X-ray interpretation

FDA approves esketamine nasal spray, the first new major depression drug in more than 30 years

A gentle method for unlocking the mysteries of the deep brain