Unsupervised clustering of multivariate circular data.

Area of Science:

• Medical Physics
• Computational Geometry
• Data Science

Background:

• Radiation therapy involves precise positioning of X-ray beams around patients.
• Patient positioning data, represented as angles on a circle, presents unique clustering challenges.
• Existing clustering algorithms like k-means are unsuitable due to data characteristics.

Purpose of the Study:

• To develop an unsupervised clustering approach for analyzing X-ray beam positions.
• To identify similarities in patient positioning for creating pre-adjustment settings.
• To improve efficiency in radiation therapy by reducing X-ray positioning time.

Main Methods:

• Defined a novel distance metric suitable for circular data (angles).
• Adapted a simulated annealing algorithm to minimize clustering distortion.
• Validated the algorithm's theoretical convergence and practical performance.

Main Results:

• Successfully clustered simulated and real-world X-ray beam positioning data.
• Demonstrated the effectiveness of the custom distance metric and clustering algorithm.
• The proposed method addresses limitations of traditional clustering techniques for angular data.

Conclusions:

• The developed unsupervised clustering method is effective for X-ray beam positioning data.
• The approach offers a pathway to creating automated pre-adjustment settings.
• This research has the potential to significantly optimize radiation therapy workflows.