Clustering of nasopharyngeal carcinoma intensity modulated radiation therapy plans based on k-means algorithm and geometrical features

Background: The design of intensity modulated radiation therapy (IMRT) plans is difficult and time-consuming. The retrieval of similar IMRT plans from the IMRT plan dataset can effectively improve the quality and efficiency of IMRT plans and automate the design of IMRT planning. However, the large IMRT plans datasets will bring inefficient retrieval result. Materials and Methods: An intensity-modulated radiation therapy (IMRT) plan clustering method based on k-means algorithm and geometrical features is proposed to improve the retrieval efficiency from the IMRT plan dataset. The proposed method could benefit future automatic IMRT planning based on prior knowledge. In this study, a collection dataset including 100 cases of nasopharyngeal carcinoma IMRT plans was employed in the clustering experiment. The geometrical features of each cluster center were used to qualitatively predict the dosimetric characteristics of organs at risk (OARs) and compared with practical results. Results: Experimental results demonstrate that the tested dataset can be well clustered using the proposed method. The predicted dosimetric characteristics of OARs for each cluster agree well with their practical results, and the difficulty of IMRT planning for each cluster can be derived. Conclusion: The proposed IMRT plan clustering method can bring great benefit to the new cases of IMRT planning.