AI helping doctors read MRI scans

Can machine learning help doctors read MRI scans more quickly and accurately?
26 January 2021

Interview with 

Andrea Rockall, Imperial College London


Artificial intelligence (AI)


MRI can be used to image different parts of the body, and can give really detailed information about tissues. But with radiologists - doctors specialising in medical imaging - in short supply and the difficulty of interpreting some of these scans, researchers are busily working at ways to harness the potential of artificial intelligence and machine learning to translate these medical images faster and more accurately. Andrea Rockall from Imperial College is one such radiologist who spends half the week in the hospital and half the week in research - currently, training machine learning algorithms to help doctors read whole body MRI scans. Katie Haylor spoke to Andrea, asking firstly about when someone might actually undergo a whole body MRI scan...

Andrea - Mainly this would be in patients with suspected or confirmed cancer where we want to identify the initial tumor and also sites where the tumour may have spread. One of the challenges we have is that patients with cancer may have multiple different types of scans, for example a CT or a PET-CT or an ultrasound. And the concept would be that a whole body MRI may be good enough to do the full staging of the patient in one sitting.

Katie - Oh, I see. So you're not having to translate between different technologies.

Andrea - Yes, obviously each of the different technologies has pros and cons, but whole body MRI may be able to replace multiple visits to the hospital and multiple scans for each individual patient because it gives such an excellent overview.

Katie - How have you actually trained the computer to read these whole body scans?
Andrea - The way we train the computer is by teaching it essentially what an expert does. So the first step for the whole body MRI machine learning training is really to demonstrate a whole number of scans in healthy volunteers that are labeled, so different organs are labeled. So the liver and the lungs and the heart for example, are labeled by the experts. And these are fed into a computer that then learns about what the normal organs look like.

We then take a set of scans in patients with cancer and effectively the expert colours in the sites of the cancer. So labels them as sites of cancer. And then these scans are then fed into the computer algorithm in order to train the computer to recognise the appearances of the cancer on the scan.
Once we've trained the algorithm to recognise and detect these areas of abnormality, we then test the output on new scans that the computer's never seen in order to see how well the algorithm is performing.

Katie - How well is it performing? Do you have enough data at the moment to be able to say?

Andrea - With our first study, which has just been completed, we found that using the AI algorithm output with an expert reader... essentially the expert reader performed approximately the same, but perhaps was very slightly quicker at doing the read. What we did find is that there was some improvement in the detection of cancer when you asked a more junior reader to look at the scans and diagnose the disease. So it may be that this technique will help people who are training or less experienced at reading whole body MRI.

Katie - So just to be clear, what you're saying is: your initial findings suggest that the algorithm does better than slightly more junior members of staff?

Andrea - Well, don't forget these algorithms are being read in the presence of this expert, so the expert has an overlay of the machine learning output; and their performance is compared to how they perform without the machine learning output compared to with. It's not the machine working on its own; the machine does not perform well enough at the moment to do that.

Katie - Okay. So it's very much a joint venture between an expert human and hopefully an expert algorithm.

Andrea - It is exactly that. The expert or the junior reader will be provided the scans, with or without the machine learning output, to see how that affects their diagnostic performance.

Katie - What is your hope? Is it that, with the help of some sort of machine learning algorithm, doctors will be able to read these body scans faster, or more accurately, or both?

Andrea - We would hope for both! Firstly, as the scans are going through the scanner, if there's a serious abnormality that needs urgent attention, we would like to see that those scans would be triaged quickly so that they come to the top of the list for the radiologist to report. Secondly, it would be great if areas of concern could be highlighted for the radiologist in order to help them with the read. But thirdly, if we can speed up the time it takes radiologists to read, that would be wonderful as well, since we're short on radiologists in the NHS.

Katie - Are there other areas of healthcare where we're already doing stuff like this?

Andrea - Yes, whole body MRI is at the very far end of the spectrum. Where we have successful AI outputs that are available for clinical care include looking at CT scans of the brain in patients with suspected stroke; so these are available now in the NHS, and here we can speed up differentiation between somebody who's had a bleed or a blood clot, and that can speed up the triage to treatment. And for example, in Accident and Emergency, there are AI tools which are approved for detection of fractures, for example; or something called a pneumothorax in the chest; or pulmonary embolism, where you get a big blood clot in the chest. And some of these are already available in much simpler technologies than the whole body MRI.

Katie - How long do you reckon it'll take before whole body MRI scanning, as read by a computer, would actually be in hospitals?

Andrea - I hope perhaps in the next five to eight years. But it takes this long just to really, really develop the technique to a stage where it is safe for application to patient care, and where we've actually tested it, a little bit like we would run a drug study to make sure that the diagnostic tool is working effectively and safely. Doctors are quite conservative, I would say, and they need pretty solid data with evidence of safety before rolling out. Some of the things in the slightly simpler technologies, like plain films or CT scans, will be coming through already now; but also more and more tools will be coming through in the next five years or so. The whole body MRI, I think, is going to take a little bit longer.


Add a comment