High hopes for AI in healthcare: ‘Now we have to prove it’

Peter van Ooijen, full professor of AI in Radiotherapy, researches how AI can contribute to the optimal treatment plan for patients. In this episode of the JTS Scholars series, he discusses his work and shares his views on the implementation of AI. 'You can compare it to the first navigation systems: in the early days, the front-seat passenger would still be looking at a printed map.'

Text: Jelle Posthuma

‘We must stop training radiologists now.’ With that statement, Geoffrey Hinton, AI pioneer and Nobel Prize winner, caused quite a stir in 2016. In his view, the introduction of artificial intelligence would make radiologists redundant. ‘In hindsight, it was a totally unfortunate statement,’ says Van Ooijen. ‘Radiologists stepped on the brakes out of fear of losing their jobs. I still remember a leading radiologist standing up at a conference years ago and saying: “How are we going to stop this?”’

Van Ooijen notes that attitudes towards AI have now clearly changed. Many radiologists no longer fear for their jobs and see artificial intelligence as a valuable tool. ‘Nowadays, you often hear people say: AI won’t replace radiologists, but radiologists who don’t use AI will eventually be overtaken by colleagues who do. Technology such as AI can be a great help to radiologists. That’s what Hinton should really have said. In some cases, AI even performs better than radiologists, but it doesn’t make them redundant.’

Predicting the future

Van Ooijen conducts research into adaptive radiotherapy using AI. This technology allows radiotherapy to be personalized to the individual patient. In current clinical radiotherapy practice, a treatment plan is drawn up in advance, after which the radiotherapy fractions take place and, subsequently, the outcome for the patient is assessed, explains Van Ooijen. ‘During treatment, both the patient and the tumour change, but the treatment plan remains the same.’

Artificial intelligence makes the treatment plan adaptive: it can be adjusted during the course of treatment. ‘For example, we use AI models for segmentation: delineating the tumour and the surrounding organs at risk.’ This time-consuming process can be automated using AI, enabling adjustments to be made between different treatment sessions. The next step is to develop a model that can predict treatment outcomes. ‘Our model predicts potential adverse effects and the likelihood of a treatment being successful. Using these predictions, we can identify the optimal treatment plan for the patient.'

The segmentation of organs at risk using AI is now widely used in radiotherapy, Van Ooijen continues. However, the use of artificial intelligence in healthcare still faces a number of challenges. According to the professor, one of the main obstacles lies with the user, who is sometimes still reluctant to embrace the technology. ‘The debate is no longer so much about job losses, but about the loss of knowledge and skills. Radiologists who currently monitor the quality of AI have carried out those tasks themselves. But that may no longer be the case for tomorrow's specialists. Will they still be able to carry out the checks?’

Road tripping without a map

Van Ooijen makes a comparison with the introduction of navigation systems. In the early days, the front-seat passenger would still have a map on their lap to check the system. ‘Nowadays, we rely on it blindly. The younger generation often doesn’t even know how to read a paper map anymore. Is it a bad thing that we’re losing that knowledge and those skills? Not always. But for certain critical skills, it’s a different story. The same questions also arise when it comes to the use of AI in healthcare.’

Another major challenge is the explainability of AI models, the professor continues. Artificial intelligence is often seen as a sort of ‘black box’ – put something in at one end and a result comes out at the other. 'Our applications therefore focus on developing explainable models that provide the user with more information, for example through uncertainty visualization. Using a probability map, our model can indicate how certain it is that a particular area of the body is part of a tumour: red means almost certainly a tumour, blue almost certainly not, and everything in between.’

Explainable AI also has implications for the legislation and regulations concerning artificial intelligence in healthcare. ‘In Europe, human oversight is a fundamental principle: in critical processes, an expert should always monitor the system to ensure accuracy. Explainable models provide the expert with a tool to carry out this monitoring effectively.’

High expectations

The professor knows that interdisciplinary collaboration is essential when introducing new technology. Van Ooijen is the former theme coordinator for Digital Healthcare and is now a Scholar at the Jantina Tammes School. According to him, the School plays an important role in bringing the various disciplines together. ‘That is crucial. The expertise is certainly present within our organisation, but people don’t always know how to find one another.’

It is not just about legal expertise, but also about issues relating to human behaviour, business administration and change management. Van Ooijen notes that exactly these challenges regarding the implementation of AI are currently a major focus in the medical world. ‘Initially, innovations were mainly focused on high-risk healthcare applications. Now, however, the focus is increasingly shifting towards low-risk areas, such as summarising patient records using language models. The key question today is: which applications do we actually want to implement? Which tasks are cost-effective for AI to carry out?’

Whether AI carries out tasks more efficiently and effectively remains to be proven in many areas, says Van Ooijen. ‘Expectations are high, but in many cases the technology is still very much in the early stages. We still have to prove it.’ The introduction of new technology often involves ‘trial and error’, but in healthcare this is difficult. 'Healthcare regulations are stricter and the requirements are higher. The process takes more time and energy.’ Van Ooijen returns to his comparison with the navigation systems. ‘In the beginning, the passenger also spent more time following along on the map. I think we now need to go through that same process with AI innovations in healthcare.’