Data mining to aid beam angle selection for intensity-modulated radiation therapy

Research on the beam angle optimization problem for intensity-modulated radiation therapy (IMRT) has focused on developing optimization approaches to obtain a single high-quality angle set for each patient. In this paper, we propose a population-based method to aid the beam angle selection process with the goal of significantly improving the efficiency of the process while identifying high-quality angle sets. We use a database of approximately 2,700 IMRT treatment plans for 10 patients with locally advanced head and neck cancer to train three data mining techniques (linear regression, neural networks, and k-nearest neighbor) to rank treatment plans based on potential quality. Using 5-fold cross-validation, we train our models to rank a list of 54 new beam angle plans (for a total of 50 lists) and then fully evaluate a set of the most promising plans. Our models can, choosing only three plans out of 54, correctly identify at least one of the top three plans up to 76% of the time. Furthermore, 34% of the time we find the top plan. By sorting the plans prior to full evaluation, we are able to find the better plans faster.