Session Item

There are many complex systems that physicists use where they are not experts in the mathematical details. Examples are dose calculations like Collapsed Cone, Acuros, plan optimization algorithms, automated optimization, the exact working of linac/magnetron, etc.

If a radiotherapy company sells a product that is based on AI/ML, all software is CE-marked and tested by other users. This, of course, doesn’t guarantee a perfect working of such software, but all software used in the radiotherapy field needs to be commissioned. Also AI/ML based software needs extensive testing in your own hospital, with your own test cases. As a starting user of a ML based technique, make sure you are not the first user and others have tested it as well. In addition, one can ask the manufacturer to provide published data of tests to get a first idea of how it works.

Of course, a few basic facts need to be understood. ML methods learn from cases. The more cases used for learning, and the more spread in cases, the better the product can be. And we have to know roughly which cases were used for training as ML can also introduce bias. When software is trained on a solely western population, there is no guarantee it would work as good on e.g. an Asian or African population; if training was done mostly on male cases, it could work less favorably for females. In addition, we have to keep in mind the limit of our own knowledge: our own contours and treatment plans are subject to a large variation, and although ML may not be perfect, it could end up with less variation than when we do everything without the support of software.