Artificial intelligence (AI) in medical imaging is a topic of much debate and insufficient consensus. To understand how radiologists think about AI, researchers have analysed their knowledge, perceptions, and discourse. Yet beyond these useful figures and conclusions, the conversation in itself is as compelling.
In this article, we give the floor to five radiologists from the Netherlands and the UK. Their views both coincide and differ, reflecting their own experience. The interviews are part of our series ‘Opening the Black Box of AI in Medical Imaging’, which you can watch here.
Workload vs hype
Prof Dr Hildo J. Lamb, MD, PhD, Director at Cardiovascular Imaging Group, starts by saying that the increase in imaging demand has reached a tipping point. Soon, radiologists will no longer be able to cope with their workload:
“For the past ten years or even longer, the average workload increase was 12-13% every year. Compared to the early 2000s, we are at 400% workload with the same number of radiologists, or maybe even fewer. (…) Therefore, we think the only solution is AI.”
Why is AI the technology to save the day, and how, exactly, does it aid radiologists? It can be hard to see through the hype, which Drs Alexander Scholtens, Radiologist at Tergooi, notices first-hand:
“AI is ever-present: everybody wants it, nobody knows precisely what it is. Everybody is looking forward to implementing it as much as possible in daily practice, but it’s still in development.”
Going beyond the hype, the radiologists continue by sharing their perspectives from actual clinical practice.
The AI buddy
To explain the value of AI in radiology, Dr Lamb describes a spectrum. At one end, the AI autopilot requires no human intervention. On the other end, the AI police check everything the physician does and punishes mistakes. The worthwhile concept to him, however, lies in the middle: the AI buddy:
“The AI buddy is a second reader that tries to prevent you from making errors. For example, when you say ‘Authorise the exam!’ in your dictation, the AI pops up and says, ‘Hey, did you see this nodule? Do you think it’s relevant?’ Like a friend, helping you.”
Marieke Vermaat, Radiologist at Canisius Wilhelmina Hospital, makes a similar point:
“I see it [AI] as a second reading, as an extra pair of eyes looking [at the scan] together with you. And I think that is a good thing.”
How does the AI extra pair of eyes improve a radiologist’s workflow – and why do physicians trust it?
What computers are great at
An AI second read complementing the radiologist’s intelligence increases confidence, a word repeatedly used by Dr Graham Robinson, Consultant Radiologist, NHSE Clinical Lead for Digital & Imaging Transformation and President of the British Society Thoracic Imaging. Referring to the lung cancer screening initiatives in the UK, where several sites use AI for lung nodule management, he says:
“The radiologist will occasionally miss a nodule. (…) I rely, to an extent, on the confidence of having the [AI] nodule detection behind me. (…) And when I detect a nodule, [the AI] provides an automated, reliable volume measurement, which is essential in the evaluation of nodules.”
Developing on his argument, Prof Lamb also highlights the benefit of AI supporting tasks humans are not so good at, such as precise measurements:
“Quantification is very difficult [for humans]. Is a nodule bigger or smaller? Is it sharper defined or less so? Has the density changed or not? And how does it compare to 10 previous scans? That can be a great area for AI solutions because it is very hard for a person to quantify it quickly.”
Dr Laurens Topff, Radiologist at the Netherlands Cancer Institute, sums up these benefits as:
“AI applications are designed to perform consistently and are reliable additions to the radiologist workflow. And they’re really good at quantifying abnormalities, whereas the radiologist is good at a more qualitative interpretation.”
“Consistent and awake”
Drs Alexander Scholtens, who has also been using AI for lung nodule management, adds another benefit to the discussion: the 24/7 high performance of an AI-powered assistant:
“[AI] is never sleepy. It is never distracted. (…) You have a very consistent and awake partner in crime. Yes, it will have false positives and false negatives, but that’s why I am there. So, the main benefit for me is an added pair of eyes on a specific subject in a very consistent way.”
He goes on to illustrate the point with an example from his practice:
“Following up [on lung nodules], it’s always hard to do the exact same measurement as the previous human. But if the same artificial solution does the measurements, you won’t have a jittery hand or a moment of distraction causing a tiny difference in caliper placement. The algorithm-based measurement, which is absolutely consistent, is more secure.”
Marieke Vermaat specifically experienced this advantage during the Covid-19 pandemic:
“Because a lot of chest CT scans were made [during the pandemic], also during the night, it [AI] was (…) a very good support.”
Balancing the limitations
Radiologists are also critical of AI in medical imaging. For starters, workflow integration is not a given. Not all available AI clinical applications provide their results in the existing radiology systems (the PACS). Seamless integration is essential to facilitate, as we explain in one of our previous articles.
Whilst Dr Lamb sees AI as the possible answer to the workload challenge, he is sceptical about its short-term time-saving benefits:
“I am afraid that in the next five years, maybe longer, AI will not save us time. It will cost us time because we have to check all the results. We have to decide: do we integrate the quantitative results in the report or not?”
Dr Robinson acknowledges the perception of AI tools adding to the radiology workload. However, he provides a possible solution:
“To me, it’s about what information you actually need [from the AI solution] in the here-and-now, to improve your workflow. (…) I think there’s a fear that AI picks everything, and you have to act on everything, but that’s not the case.
For example, doing a CTPA, I may just want to have the comfort of a background AI read for pulmonary embolism. The software could do lots of other things, but you can turn those off, so you answer that one question.”
Dr Topff’s hesitation, on the other hand, is that the information provided by the AI may be insufficient for the complex interpretation he performs in clinical practice:
“[AI] focuses on really specific narrow use cases, which is in contrast to the radiologist who has to make a lot of interpretations at the same time.”
No AI without the radiologist
None of the radiologists interviewed fear being replaced by AI tools. As Dr Vermaat puts it:
“A computer system is very good at standardised reporting and assessment, but the nuance and seeing in a clinical context, that is our job”.
Nonetheless, the collaboration between the radiologists and AI is a successful one that positively impacts the patient, as Dr Topff explains:
“I think the AI systems of today are really good at complementing the radiologist, augmenting the radiologists, and that way increasing or just improving patient care.”
Moreover, Drs Scholtens looks forward to using more AI applications and doing his job even better:
“I hope we’ll be able to implement many useful AI solutions in the coming years. They can help us provide adequate patient care and the best possible data reading for all clinicians who need us to provide good answers.”