Artificial intelligence helps deciding on best radiation therapy for prostate tumours

The more targeted you can radiate tumours, the better. After all, you want to spare healthy tissues and organs around the tumour as much as possible. In prostate tumours, this is often done with brachytherapy: several hollow tubes are placed as close to the tumour as possible, after which a radioactive source is passed through them.

The biggest challenge here is determining how long the radiation source should stay in those tubes. This determines how much radiation is delivered there. If it is too little, you won’t get rid of the tumour properly, but if you leave the source in place for too long, you damage healthy tissue.

Artificial intelligence

It takes a lot of time to determine the right duration for each position. This varies from patient to patient. But since the treatment plan is created only after the hollow tubes are placed, these calculations should not take too long. Typically, practitioners stopped manually adjusting the treatment plan after 30 to 60 minutes, even though they still did not know whether they had the best plan in hand.

Doctors at Amsterdam UMC felt that this process can be improved. In collaboration with researchers at the Centrum Wiskunde & Informatica (CWI) and company partner Elekta, an algorithm was developed, with accompanying software. This artificial intelligence (AI) provides a number of treatment plans that suggest the best trade-offs between irradiating the tumour and sparing healthy tissue. The doctor can choose which plan best suits the patient.

Only 30 seconds

The algorithm used is specially designed for complex problems with thousands of variables. “Normally, an algorithm has no further information about the problem it has to find a solution for – the problem is considered a ‘black box’. But as a result, the optimisation process takes a very long time,” explains Anton Bouter, who is currently a scientific software developer at CWI’s Evolutionary Intelligence (EI) group.

“Our form of AI is an evolutionary algorithm that uses information about how the problem is put together.” The problem must be formulated in such a way that the AI comes up with the best solution. Bouter co-wrote the formulation, but with that he and his colleagues were far from there. “If you included the complexity of the problem in your calculations, you did get high-quality plans, but that took the AI an hour. That takes too long for the patient. So we had to come up with something that would still give you a good plan in a short time.”

Bouter managed to achieve that by using certain parallelisation techniques. These ensured that the AI takes not an hour, but only 30 seconds, to come up with plans that are acceptable to the attending physician.

“The biggest challenge in my research was getting the most out of the algorithm. Many calculations have to be done for each organ to determine the right dose of radiation – we are talking about tens of thousands of points. These calculations have to be done simultaneously to save time. For that acceleration, I used the parallelisation techniques.”

The best choice

The beauty, says Bouter, is that physicians do not get one treatment plan, but a number of plans that are optimal for their patient. Each with a different trade-off: in one plan, the tumour receives a higher dose of radiation than in another, resulting in more or less damage to certain organs. Based on the patient’s age and condition, the doctor can make the best choice from these. “The doctor remains ultimately responsible for the treatment, not the algorithm or its creators.”

Other applications

Bouter worked on a simple interface for doctors that works well with existing clinical software. That interface is still being refined. Amsterdam UMC is already using the algorithm and associated software, called BRIGHT. Meanwhile, BRIGHT’s radiation plans have been shown to be better than doctors’ calculations.

In theory, the optimisation method has many more applications. For instance, a national study is under way to see if the algorithm is useful in the treatment of cervical cancer. Bouter also investigated his optimisation methods for medical image registration (MRI or CT scans).

*Optimal mixing evolutionary algorithms for large-scale real-valued optimisation. Including real-world medical applications. Anton Bouter, 13 February 2023. Supervisor: Prof Dr Peter Bosman. Co-supervisor: Dr Tanja Alderliesten

(This message has been reproduced from CWI.nl)