Performance Evaluation of Deep Learning Architectures for Predicting 3D Dose Distributions in Automatic Radiotherapy Treatment Planning

Radiation therapy is a leading treatment methods for cancer today, but its advancement has come at the cost of increased complexities of treatment plans, and increased planning time. The planning technique involves optimizing each plan through repeated trial and error to meet the Organs at Risk (OAR) dose criteria. This paper is a comparative study, analyzing the performance of two deep learning architectures, U Net and Generative Adversarial Network (GAN) in determining the radiotherapy dose distributions received by the tumour and nearby organs. The Dataset is taken from the OpenKBP 2020 AAPM Grand Challenge hosted by CodaLab and consists of 3D head and neck CT scans of patients diagnosed with oropharyngeal cancer. To evaluate the models we use dose scores and DVH scores, evaluation metrics proposed by the OpenKBP challenge. The Dose Volume histogram (DVH) is also used to compare the treatment plans. The dose error for the U Net model was calculated as 3.798 and the DVH error was calculated as 2.732. For the GAN model, we achieved a dose score of 3.850 and a DVH score of 3.215. The U Net model outperformed the GAN in predicting the dose distributions.