Over 60,000 patients are diagnosed with Cancer in Sweden every year, and about 300-400 of these are children. About half of all cancer patients in Sweden undergo radiation treatment. Before the radiation treatment begins, an oncologist must manually draw a map of the parts of the body that will be treated by radiation, and the regions that are important to avoid.
Manual segmentation is however time-consuming and challenging for the oncologist. There are software to automate this work, but current methods are either slow, inaccurate for patients with atypical anatomies, requires larger amounts of data than what is typically available, or are only developed for adults. Methods based on the latest developments in artificial intelligence (AI) have gone through a fast-paced development in the last years, and work impressively–on adult patients. Corresponding tools, that also work for paediatric cancer patients, are however severely lagging.
The purpose of the proposed project is therefore to develop and evaluate modern AI methods, based on deep learning and deep neural networks for segmentation in medical images, and that also work on paediatric cancer patients. We will develop a protocol for evaluating automatic segmentation methods in the clinic, at the University Hospital of Umeå. We will evaluate current state-of-the-art segmentation methods, developed with images from adult patients, on data from paediatric cancer patients, to quantify the state of the problem for paediatric cancer patients today. Finally, we will develop novel segmentation methods, that guarantee the same results for both adults and children.
The proposed project will take a first step towards a Swedish standard for automated radiation treatment planning that includes all patient groups, and facilitate the use of such AI methodology in clinical practice in Sweden.