Automated Learning of Radiation Therapy Dose Distribution for Head and Neck Cancer

Davood Hajinezhad, Afshin Oroojlooy, xin Hunt, MohammadReza Nazari, Jorge G. Silva, Shiva Das

Keywords: Radiotherapy, Machine Learning, U-Net

Abstract:

In the past decades, 3D dose optimization, with the goal of providing a prescription dose to the tumor while maximally sparing surrounding healthy organs at risk (OARs), has been playing a pivotal role in external beam radiation therapy. To reach a practical clinical treatment plan, the optimization solution needs to be adjusted by human planners. However, this procedure necessitates intensive computations combined with iterative inputs from an expert human planner. This is a time-intensive process with dose distribution solutions that are highly dependent on the human expertise of an institution.
To remedy the computational time issue, and also reduce the variability, while capturing the existing knowledge, we use machine learning capabilities. In particular, we propose a suitable feature matrix for head and neck cancer and train Deep Neural Network models to mimic the Planned Dose Distribution procedure. When trained, the model can provide dose values in less than a second. Toward this, a specific Convolutional Neural Network called U-Net model is utilized. The numerical results show that this model reduces the solution time of each patient on the utilized data from 3-6 hours to 1 second while providing highly accurate solutions such that the average error is approximately 3.2\% and 7.0\% for cancerous organs and OARs, respectively.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: both

Primary Subject Area: Application: Radiology

Secondary Subject Area: Integration of Imaging and Clinical Data