Professor Anwar Padhani, our lead consultant for MRI, talks about his work to pioneer artificial intelligence software in cancer imaging at Paul Strickland Scanner Centre.
It seems like every day now that we hear of new advances in artificial intelligence (AI), a development that promises to change how we live and work forever.
Paul Strickland Scanner Centre is among a select groups of cancer imaging centres where game-changing artificial intelligence software is being initially deployed.
World-renowned cancer imaging expert Prof Anwar Padhani, our lead consultant for MRI, is pioneering the use of AI software and spoke to Inside View about the promise it holds for finding prostate cancer.
Prof Padhani said: “We’re currently trying out different sequences to see where AI acceleration would be best at the image acquisition phase.” Not only can AI serve as a second pair of eyes, but it can save radiologists time by automating some of the routine tasks involved in analysing a scan, allowing a doctor to be more productive.
Some of the ways in which AI can speed things up include outlining the margins of tumours or counting and measuring the number of lesions in a scan, which can be a laborious and timeconsuming task for a radiologist.
‘Some AIs have already reached expert level’
This can shave several minutes off the time it takes for a scan to be analysed, which could lead to a radiologist reporting more scans during their shift than otherwise, thereby increasing the number of patients a scanning centre can serve over time.
In addition, AI technology can be used to make images clearer, something called “artefact” or “noise” reduction in radiology, which can lead to a more accurate diagnosis. The AI can also act as a second reader after a radiologist has analysed a scan. However, relying purely on artificial intelligence technology to detect and interpret scans is still some time off, according to Prof Padhani.
“Even though it has learned from thousands of scans, the AI misses cancers and a radiologist will therefore still need to check its work. “For example, the AI will sometimes tell you that something is not a cancer – then the radiologist would contradict it and say, ‘that it is actually a cancer’. This is how the programme learns and improves.
“Over time, human readers get tired and can make mistakes, particularly for repetitive tasks. “AIs don’t get tired and are great at repetitive tasks.” Prof Padhani explains it is therefore vital that the AI is only trained by the very best experts in their field. “If there are non-experts amongst the group of radiologists who train the AI, then the AI might not spot cancerous lesions due to having taken on board the skills of the non-experts.
“There will be a degradation of AI performance if you let the AI learn from everyone it is exposed to. The obvious answer is to only let the AI learn from experts, but experts are expensive and their time is very limited.”
Training the AI can be a time consuming, painstaking and extremely boring task for an expert radiologist, which means it is difficult to get scarce and already very highly paid experts to commit to the task.”
With regards to prostate MRI for detecting early cancers, he said: “When it is trained by experts, the AI improves its abilities very quickly and rapidly levels up from generalist to specialist. Radiology AI software that is currently used as a diagnostic aid is as good an average radiologist, with similar levels of false positives and false assurances, however I have recently seen data which suggest that some AIs have already reached expert level.”
“It’s not dissimilar to human training, actually. “The better your teachers are, the better your learning will be. Similarly, there is a point at which the improvement shown by the AI levels off. Just like someone who has been a cancer imaging expert for so many years, it takes a lot of extra training to improve someone who is already at the top of their field to improve even further.
Prof Padhani believes that in the future AI could not just help radiologists analyse images but even write up the radiologist’s report on the scan, which could be used by the referrer to make treatment decisions.
“I think we will get there. It will emerge in limited circumstances at first. Let’s say that one day you deployed an AI in a prostate cancer screening programme. In such a scenario, the decision is a very binary one – it’s not ‘does this patient have cancer or not’ but ‘should we, or shouldn’t we call the patient back’ for further checks. People don’t usually get a letter in the post that says they have cancer – they are invited back for further
assessment.”
This could be an easy task for an AI, provided a suitable model has been developed. “The scan and the accompanying report would raise a flag that something may be going on that needs further investigation. That would help to reduce the harm from false alarms. “There is no doubt that the AI is improving as a diagnostic aid, although we cannot yet rely on it as a decision support tool. We cannot base clinical decisions on what the AI tells us at present.”
Prof Padhani hints at what he thinks the future might hold. In the past 5 years or so, the software has made tremendous progress, and with data indicating that it is already surpassing the abilities of an expert, Prof Padhani believes the future could look very different to the present. “The tools that are now being built look very promising.”
Each improvement in the technology builds on the last one, which means that advances compound and can have an astonishing effect over time. “A generalizable, decision support tool will require studies in multiple centres on different machines against a stronger gold standard; these studies are only now beginning. Ten years from now, we will be in a completely different place.”
Prof Padhani believes that AI will be particularly useful if a largescale prostate cancer screening programme using MRI scans were ever put in place.
“We simply do not have enough radiologists currently to report all those scans – and the shortage of staff is set to become worse in the future and not better due to too few radiologists having been trained in recent years to meet demand.”
“AI assistance will be required when deploying a community-wide, MRI-driven, prostate cancer diagnosis pathway for suspected and screen-detected men,” he concluded.
