AutoML framework for dose prediction models in radiotherapy

Radiotherapy is one of the main types of treatment against cancer. The aims of any radiotherapy is to destroy the tumorous cell while sparing as much as possible the healthy tissues. When a patient needs radiotherapy, a plan must be established. To do so, one must first get the images of the patient, then physicians must contour the tumor and the different organs at risks, after that some thresholds have to be set for each organ and for the tumor, finally a dosimetrist have to generate the plan. This process takes many hours and computational resources. Artificial intelligence is nowadays used in these different steps in order to speed it up but also, reduce the amount of work needed from the specialists. Automated machine learning (AutoML) is a new trend in machine and deep learning world. Its goal is to propose an easy way of making model selection and parameters tuning for any dataset. Thus, reducing the amount of time spent at manually doing these steps. A secondary goal is to have more people being able to use such software even if they are not specialists in the domain because the model will do the complicated steps by itself. This thesis wants to reunite both worlds, by proposing an AutoML framework for the dose prediction problem in radiotherapy. It is based on previous work on HD-Unet and adapt it to use the principle of AutoML. It is easy to use, maintain and evolve. For a new dataset, it computes many parameters without the need for the user to do anything. To test this model, multiple datasets were used to show the adaptation of the model. Its results are better than the work it is based on concerning H&N cancer and even if not as good as the fine-tuned model on other datasets its results are still interesting.