Every year, about 35 children in the Netherlands are diagnosed with a kidney tumor. Their treatment typically includes chemotherapy, surgery, and can include radiotherapy. While radiotherapy plays a crucial role in eliminating remaining cancer cells, it can also harm healthy organs nearby.
Helping children receive safer cancer treatment
Since 2015, in Utrecht, flank irradiation using highly conformal target volumes has been adopted. This method adjusts for the way organs can shift after surgery, allowing for more precise targeting of the radiation dose. Dr. Geert Janssens is a radiation oncologist at the Princess Máxima Center and UMC Utrecht: ‘To make this work, we must outline the various organs and structures at risk near the radiation area, or so called target volume, on every slide of the scan. Doing this by hand takes a lot of time and the quality of contouring depends heavily on the experience of the team.’
Building an AI tool for children’s anatomy
Existing AI tools to outline the organs and structures at risk, like the liver, spleen, heart and bowel, work well for adults, but they often fall short in children. That’s because children are still growing. Their organs vary significantly in size, shape, and less fatty tissue is surrounding them, so an AI model tailored specifically to children's anatomy is needed.
Mianyong Ding, a PhD student in the Van den Heuvel-Eibrink group and Butterfly student, created a deep learning model specifically for radiotherapy in children with upper abdominal cancers. He worked under the guidance of Dr. Janssens, Dr. Matteo Maspero, assistant professor in the computational imaging group and medical physicist resident at UMC Utrecht, and Prof. Marry van den Heuvel-Eibrink, research group leader and pediatric oncologist at the Máxima Center.
Ding trained the model using 189 CT scans from children treated at the Princess Máxima Center and UMC Utrecht, and another 189 publicly available pediatric scans. As a result, the model can automatically detect and segment 17 different abdominal organs and anatomical structures, such as the liver, pancreas, and spleen.
‘Our model delivers accurate results on both our in-house dataset and a public dataset’, says Ding. Van den Heuvel-Eibrink adds: ‘We also tested it in a workshop where we asked a panel of international pediatric radiation oncologists to rate the model-generated organ contours on a scale from 1 to 5, with most ratings indicating clinical usability.’
The researchers conclude that the model is ready to support clinical settings. The study’s results were published in Radiotherapy and Oncology. The model will be implemented within UMC Utrecht's framework to allow deployment in clinical practice where children of the Máxima Center receive radiotherapy.
Improving radiotherapy quality worldwide
By automating the contouring of organs and structures at risk, the AI model saves time, and reduces variability between clinicians, based on findings from an earlier follow-up workshop.
‘This tool brings more consistency and quality to pediatric radiotherapy in particular for centers with less experience,’ says Janssens. ‘And because we’ve made the model open source, other hospitals can benefit too.’
The Princess Máxima Center is now working with KiTZ Heidelberg to validate the model's robustness and ensure its reliability across different hospitals using different scanners and scan techniques. As a next step the researchers will focus on developing a tool for auto-contouring of the flank target volume in children with Wilms tumors.
The study was made possible thanks to the partnership between the Máxima and the Hopp-KiTZ pediatric tumor center in Heidelberg, Germany that is made possible financially in part through the Princess Máxima Center Foundation and the Horizon Europe/Marie Sklodowska-Curie COFUND project number 101081481, the Princess Máxima Center and the participating research groups and UMC Utrecht.