Open AccessTechnical Note: Using k‐means clustering to determine the number and position of isocenters in MLC ‐based multiple target intracranial radiosurgery
Author(s) -
Yock Adam D.,
Kim GweYa
Publication year - 2017
Publication title -
journal of applied clinical medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.83
H-Index - 48
ISSN - 1526-9914
DOI - 10.1002/acm2.12139
Subject(s) - isocenter , radiosurgery , cluster analysis , centroid , metric (unit) , mathematics , algorithm , k means clustering , wilcoxon signed rank test , position (finance) , explained sum of squares , radiation treatment planning , computer science , statistics , nuclear medicine , medicine , radiation therapy , geometry , radiology , operations management , mann–whitney u test , finance , imaging phantom , economics
Abstract Purpose To present the k‐means clustering algorithm as a tool to address treatment planning considerations characteristic of stereotactic radiosurgery using a single isocenter for multiple targets. Methods For 30 patients treated with stereotactic radiosurgery for multiple brain metastases, the geometric centroids and radii of each met were determined from the treatment planning system. In‐house software used this as well as weighted and unweighted versions of the k‐means clustering algorithm to group the targets to be treated with a single isocenter, and to position each isocenter. The algorithm results were evaluated using within‐cluster sum of squares as well as a minimum target coverage metric that considered the effect of target size. Both versions of the algorithm were applied to an example patient to demonstrate the prospective determination of the appropriate number and location of isocenters. Results Both weighted and unweighted versions of the k‐means algorithm were applied successfully to determine the number and position of isocenters. Comparing the two, both the within‐cluster sum of squares metric and the minimum target coverage metric resulting from the unweighted version were less than those from the weighted version. The average magnitudes of the differences were small (−0.2 cm 2 and 0.1% for the within cluster sum of squares and minimum target coverage, respectively) but statistically significant (Wilcoxon signed‐rank test, P < 0.01). Conclusions The differences between the versions of the k‐means clustering algorithm represented an advantage of the unweighted version for the within‐cluster sum of squares metric, and an advantage of the weighted version for the minimum target coverage metric. While additional treatment planning considerations have a large influence on the final treatment plan quality, both versions of the k‐means algorithm provide automatic, consistent, quantitative, and objective solutions to the tasks associated with SRS treatment planning using a single isocenter for multiple targets.