Does AI herald the end of the radiologist?

Key insights

• Radiology leads the way in the use of artificial intelligence.
• AI is particularly suited to supporting the radiologist's core tasks but can be used in any step of the workflow.
• AI’s success hangs on its seamless integration in the existing workflow and demonstrable improvements in efficiency and treatment outcomes.

AI is not a new phenomenon but goes back to the 1950s. Machine-learning¹ made a commercial breakthrough in the 1980s and 1990s, but only gained momentum in the past five to ten years thanks to the exponential increase in computing power, big data, and the development of cloud technology and neural networks. One breakthrough was the deep-learning algorithm AlexNet. In 2012, it won the ImageNet Challenge with significantly lower error rates than had previously been possible. ImageNet is a database containing millions of hand-annotated images of animals, plants, and objects, which computer algorithms have to recognise as many as possible of during an annual competition.

Shifts

We are currently experiencing a fourth industrial revolution, which is noticeable in medicine in general and radiology in particular. Thanks to data integration capabilities, we can see a shift towards personalised medicine. Data quantification and automated image analysis are gaining in importance, and there is a growing requirement for automation. Not only because of the ever-increasing workload, radiologists tasks are also becoming more complex and medical imaging is growing in importance in therapy. Radiologists are increasingly involved in multidisciplinary teams.

Radiology leads the way

It should come as no surprise that, today, radiology leads the way in the area of AI and, in particular, machine-learning and deep-learning applications. After all, AI is ideal for image classification and analysis, which is precisely the core task of radiology. Moreover, the specialism has extensive categorised datasets and has a large market. In 2016, 23 million CT/MRI scans were carried out in Germany. In Belgium and the Netherlands, this was still up to ten times less, but this market is also growing rapidly.

Future potential

Most importantly, AI has a vital role to play in supporting the core tasks of detection, characterisation, and monitoring of abnormalities and injuries on scans. The algorithm will usually help the radiologist with image interpretation; in some cases, it can take over the radiologist's task entirely and, for example, send a CT scan directly to the emergency doctor. NKI is currently testing an algorithm that can fully automatically analyse and comment on chest X-rays. If it is not possible to view scans immediately due to a high workload, AI can help to determine priorities by automatically identifying abnormal scans, such as scans with pulmonary embolisms² or abnormal mammograms. Also promising is the use of AI for multimodal integration, i.e. the combination of data from different sources, such as various radiological examinations, pathology results, surgical information, and other patient data. The ultimate goal is to be able to make more accurate predictions of the treatment effect, for example in cancer.

More than just image analysis

Although image analysis is a vital task, the radiologist's workflow amounts to a lot more. In principle, it is possible to use AI in all steps of the workflow to make it more efficient and improve the service for the patient, from management of appointments to display protocols, invoicing, and adapted communication.

From threat to opportunity

Geoffrey Hinton, the godfather of neural networks and also involved in the development of AlexNet, caused a great deal of controversy in 2016 by stating that we should stop training radiologists now³. Such statements are grist to the mills of sceptics and may explain much of the initial resistance to AI. This resistance has now subsided, and most radiologists now see AI more as an opportunity rather than a threat. The first FDA-approved algorithms are being used in clinical practice and research, while others are being tested.

What about the patient?

However, patients still seem to be unconvinced. They will only accept AI if the physician makes the final decision, even if the AI algorithm’s diagnosis is more accurate, according to a recent study published in the Journal of Consumer Research.

Still not redundant

The chance that radiologists will soon be redundant is minimal. As yet, AI can only answer simple, well-defined questions. For example, an application that detects and monitors lung tumours can tell what type of tumour is involved and whether it has grown, but that same application cannot yet identify the collapsed lung that the radiologist would be able to see on the same scan. It will be quite some time before we get that far. In any case, developing such an application would require a massive amount of annotated training data.

Ifs and buts

So, what are the main issues and pitfalls?

• The biggest problem by far is the lack of data. Hospitals do not want to exchange data for fear of violating privacy rules. It would be a mistake to think that it is possible to resolve this shortcoming by sourcing data from countries with less strict rules, such as China. In particular for epidemiological research, it is essential to have local data to find connections between disease, treatment, and factors in the patient's environment. Moreover, annotating the data takes a lot of time and the motto “garbage in, garbage out” also applies here.
• Another thorny issue is the lack of studies that validate the diagnostic accuracy of AI algorithms in clinical practice, according to a meta-analysis published in  The Lancet Digital Health .
• Although, in principle, AI offers considerable opportunities for improving the efficiency and accuracy of radiology, we should not take its implementation for granted. Radiologists will not want to use algorithms that yield too many false-positive and/or false-negative results. The same is true if it takes too many clicks to view the results. AI’s success hangs on its seamless integration in the existing workflow and demonstrable improvements in efficiency and treatment outcomes.
• Finally, it is essential to keep the ethical aspects in mind as well2.

What is the next step?

What is required to be able to exploit AI's capabilities to the full?

• Most importantly, we need to address the lack of data. There is a need for a medical ImageNet.
• Moreover, it is necessary to improve data integration and encourage cooperation between hospitals, not only to share data but also to exchange algorithms and experiences.
• Then there is the question of AI’s clinical relevance; in other words, how do you measure treatment outcomes?
• It is also necessary to think carefully about how and when AI should be included in the training. Some fear that the position of radiologists will be downgraded, but, as Erik stressed, radiologists should not follow AI algorithms blindly. In the same way as they need to master the basics of magnetic resonance to be aware of the image artefacts that an MRI can produce, they will also need to master the basics of AI. In this context, consideration should be given to the subject of certification.
• According to Erik, there is a particular need for ‘local champions’ – AI advocates who will get behind projects – and inspiring examples of successful initiatives.
• Finally, there will have to be a killer app.

Should radiologists fear for their jobs? According to Erik, not at all. However, the role needs to be reviewed. How can radiologists use AI to give better information to colleagues from other disciplines? Because, as Dr Curtis Langlotz of Stanford University said: “Artificial Intelligence will not replace radiologists. Yet, those radiologists who use AI will replace the ones who don’t.”

1 Machine-learning is part of artificial intelligence. Unlike conventional programming, machine-learning algorithms will not tell computers how to solve problems but will teach them how to solve problems by training them to recognise patterns or rules in the data they have received. Deep-learning is a specific discipline of machine-learning that uses deep neural networks. 'The Ultimate Guide to AI in Radiology' provides more background information about artificial intelligence, machine-learning, deep-learning, and their uses in radiology.
2 A pulmonary embolism is a blood clot that shuts off the pulmonary artery.
3 The Radiology Journal published an article that provides a well-substantiated position of leading trade unions from Europe and North America.