MSc Thesis Defense by Rasmus Willigenburg Andersen


Predicting Radiography Dose Plans using Deep Convolutional Neural Networks


Radiotherapy is a normal treatment for patients diagnosed with cancer, and new radiotherapy methods are constantly being developed. Dosimetrists create dose plans to ensure that as many of the cancer cells as possible are hit by the radiation, while also ensuring that as few normal cells as possible are hit. Creating dose plans will always be prone to human errors, and when working with new methods it is not always known how much the radiotherapy will damage the healthy tissue. In this thesis, we proposed a deep convolutional neural network to predict radiotherapy dose plans from post-radiation follow up scans, and look into the damage done by radiotherapy. We evaluate the model's performance on a small set of brain tumor patient images. The results of our experiments are found to be significantly worse than the typical results of state-of-the-art methods within image synthesis. The quantitative and qualitative results showed that the model produces inaccurate dose plans and synthetic CT images. Our results shows that the model is not able to compete with state-of-the-art image synthesis methods. This is most likely due to a small dataset and the models inability to generalize. Further validation on a larger dataset is warranted, as is expansions of the network to include batch normalization, dropout and deconvolution.

Supervisor: Aasa Feragen and Sune Darkner, DIKU

Censor: Rasmus Poulsen, DTU Compute