ENGINEERINGNET.BE – When drawing up treatment plans for breast cancer, there is still a lot of manual work involved in making precise radiation plans and drawing in the organs of individual patients.
“We also simulate in a kind of flight simulator how we want to irradiate: from which side and with what intensity, so that as much radiation as possible enters the tumor and as little as possible around it,” says Coen Hurkmans, clinical physicist at Catharina.
The simulator contains a model of the treatment equipment and images of the anatomy of the individual patient. An experienced lab technician then makes a radiation plan.
That work can be automated. AI is a practical and efficient way to record, calculate and interpret knowledge and experience. Put dozens of previously made plans in a model and train the computer to make them yourself.
PDeng student Nienke Bakx of Eindhoven University of Technology assessed two software models for this purpose: the open source U-Net model, which is based on convolutional neural networks, CNNs. And the cARF model, where cARF stands for contextual Atlas Regression Forest.
Bakx filled both models with data from more than a hundred patient treatment plans. The trained models were tested during a clinical pilot with data from twenty new patients.
Radiation therapists and lab technicians made their plans for these patients by hand and compared them with the automatically generated plans to further optimize the software. The better the input, the better the output.
95% of what the computer came up with turned out to be usable without any manual adjustment. The U-Net model scored slightly better than the cARF model, which prompted Swedish partner RaySearch to use this model in the RayStation software for radiotherapy patient planning.
A fully trained, validated working model will go live in clinical practice from May. Incidentally, the radiotherapists and laboratory technicians check everything that is set up automatically.
The model is also ready for the automatic drawing of organs and glandular areas and a clinical pilot is now underway.
In the figure above: two axial cross-sections of a patient with breast cancer, showing the circumscribed breast, lymph nodes, lungs and heart.