Purpose: The quality of a radiation therapy treatment plan relies heavily on setting obtainable clinical objectives informed primarily by the experience level of the attending care team. This work aims to create data sets suitable for applications in machine learning (ML) to aid in plan quality analysis using generalized, open-source methods with the intent of providing site- and practice-specific recommendations.
Methods: Treatment planning data was collected from 40 head and neck (HN) cancer patients treated with TomoTherapy. For each patient, the dose-volume histogram (DVH) data for two regions at risk (parotids and submandibular glands (SMGs)) were obtained. Data for the parotids and SMGs were grouped into a single, bilateral set for each gland. The data were analyzed based on visualized categorization. Analysis was then repeated with the data reclassified by k-means clustering. The optimal number of clusters for each organ was explored. The average DVH and confidence intervals from visual parsing were compared to the unsupervised k-means clustering.
Results: Three distinct groups were visually observed from DVH data for parotids and SMGs, and assigned as D1, D2, and D3. The clusters determined through k-means using the organâ€™s mean dose and dose to 2% organ volume also appeared to minimize residual error using 3 centers for the parotid and SMG. The averages were recalculated using the k-means categorization and confidence intervals were improved in both size and smoothness.
Conclusion: Unsupervised clustering of readily available DVH data offers a way to parse clinical data into useful classifications. These datasets can become inputs to ML algorithms for more robust and rigorous plan quality evaluation. Future work will focus on integrating diagnosis and patient anatomy with DVH analysis in order to predict achievable dosimetric endpoints.
Funding Support, Disclosures, and Conflict of Interest: This project was supported by the Specialized Program of Research Excellence (SPORE) program, through the NIH National Institute for Dental and Craniofacial Research and National Cancer Institute (NCI), grant P50DE026787. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.
Not Applicable / None Entered.