Deep learning to predict optimized dose and its uncertainty for radiation oncology treatment planning

In recent years, deep learning has seen its potential significantly increased in the medical field, with various applications making the work of physicians more automated and efficient. One area of application where neural networks have a crucial role to play is in radiotherapy, one of the most common cancer treatments. Every radiotherapy treatment is preceded by a treatment planning phase in which an important step is to optimise the dose of radiation received by the patient. This time-consuming task could be automated using a convolutional neural network based on the patient's anatomical parameters. In this master thesis, we will use the well-known U-Net architecture for this task. However, this model would be used as a black box predicting the dose distribution with no information on the certainty of the model on its prediction. In this work, we propose to accompany these models with a map indicating, at each voxel, the level of uncertainty of the model. To do this, a Bayesian model is too costly. This is why we developed three simple methods that can be easily adapted to other architectures: the Monte-Carlo dropout, the bootstrap method and a modification of the U-Net. These methods were tested on a database of 200 patients suffering from head and neck cancer. A correlation between our measure of uncertainty and the error made by the model of about 0.27 was found on the planning target volume (PTV) and 0.56 outside. A theoretical correlation of 0.71 was computed outside the PTV.