AI tech speeds up cancer treatment

A new type of artificial-intelligence technology that can cut the time cancer patients must wait before starting radiotherapy has been pioneered by Cambridge University researchers. The technology, which will now be offered at cost price to all NHS trusts in England, has been trained to recognise and highlight the healthy tissues on a patient’s body scan that the X-ray radiotherapy dose needs to avoid hitting. This reduces the time it takes cancer specialists to map out a patient’s radiotherapy regime, dramatically speeding up treatment. Addenbrooke’s Hospital consultant Raj Jena has been leading the work…

Raj - The oncologist is trying to target the cancer within the body. Sadly, for many of the cancers that we're trying to treat, they're located deep in the bodies. The only way for an x-ray beam to get there is to go through some healthy tissue. We could get rid of almost any cancer in the body with x-rays if we could get enough dose in. The problem is you have to go through healthy tissues to get there, and that limits how much you can give. So it's as much of a task to map out the cancer as the target as it is to map out all of the healthy tissues in the path so that we can actually get the beams in a safe fashion.

Chris - What you're saying then is, because we know there are things in the way of where you want to go, you're looking for the safest route to get the maximum x-ray dose into the cancer with the least dose into the healthy tissue, so you minimise the harm. We do

Raj - Exactly right. So the robotics of the system is very, very good at arranging different radiation beams to come in from different directions. But all that's no good if we don't have a map that shows us where we need to avoid, where the no-go zones are.

Chris - And that's what hitherto has been a human very labour intensive task.

Raj - Exactly. So for an oncologist working on this, depending on where the tumour is, it can take anything between 25 minutes to upwards of two hours to mark out all of these healthy tissues. So we were looking to try and accelerate that with the AI.

Chris - How does it work?

Raj - Well, what happens is once the patient is scanned, the data goes off through a safe mechanism so that we can run our machine learning algorithm. And what's returned to the oncologist is a scan, but with extra smarts because instead of just having the image to look at, they get the image plus already we have marked out all of the healthy tissues, and that means that overall the oncologist using this technology can go about two and a half times faster. That's what we see as a sort of real world acceleration by introducing the technology.

Chris - How does the AI do that? How does it know what is healthy tissue in the first place?

Raj - So it has to be trained just as we all have to be trained. So we had to build a dataset in each case of about 150 patients. And for each of those patients, we had experts mark out exactly where the healthy tissues were. And then it took hundreds of hours to train the best model that we could. And then once we did that, we started evaluating its performance. When we got to the point where we were seeing performance that was starting to match human performance, then we knew we were onto something.

Chris - I was going to say, having built the system, did you then give it an exam to do as it were, compare it with you versus it to see if it can do a better job than you can?

Raj - Exactly. So what we did, we gave our oncologists the preparation work. In some cases it was done by the AI. In some cases it was done by their colleagues, and we didn't tell them which was which. And we actually found that in two thirds of the cases, they actually preferred to start with the AI rather than their own consultant colleagues. And that's when we really knew we were onto something.

Chris - So it produces a scan with a lot of the markup done already to guide people in the right direction. So what does the oncologist add? Is it just a safety check, or is it that there's still work to be done by the human here?

Raj - So the system hasn't learned to mark up a tumour. Tumours are much more complex and varied. Our normal tissues follow a very set pattern, so it's quite easy to get them to learn that, but it's an order of magnitude more difficult for them to learn how to segment a tumour. So at the moment, what the oncologist does is that they go straight to the tumour and they devote the lion's share of their time, mocking that out as precisely as possible. And then they have to check everything that the AI does because at the moment that is our sort of, you know, safety check, is that everything that the AI produces can't be accepted into our clinical system until the oncologist approves it and says, 'yeah, this is safe to use.'

Chris - And what sort of a difference is this making?

Raj - Well, we have introduced this and other technologies into our workflow at Addenbrooke's. Where nationally we have a 31 day target between being told we are going to go for a radiotherapy plan and actually starting it. In Addenbrooke's, for the fastest growing tumours, we're aiming for 14 days and actually we're aiming to go even faster than that. We'd like to get it down to five days if possible. And it's these sorts of technologies that help us do that.

Chris - And does this make a difference to the outcome for the patient? I know that it's a bit less time and that's maybe good psychologically that something's being done a bit sooner, but does it make a difference to disease and clinical outcomes?

Raj - Indeed, it does. What we know is that for the very fastest growing tumours that we deal with, you are 2% more likely to control a tumour every day that you can shave off that waiting time. So it really does make a difference. And on top of what you mentioned, just that feeling of sort of, you know, staring down the barrel of a gun when you're waiting and you know that you need to start radiotherapy. And you would think, 'why can't I start now?' And, that must be a terrible feeling too.