Premium
Sci—Fri AM(2): Brachy—09: Using Bayesian Networks for Prostate Brachytherapy Inverse Planning
Author(s) -
Chng N,
Spadinger I,
Salcudean T
Publication year - 2009
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.3244201
Subject(s) - brachytherapy , workflow , computer science , workload , consistency (knowledge bases) , radiation treatment planning , set (abstract data type) , prostate brachytherapy , algorithm , medical physics , data mining , artificial intelligence , medicine , database , radiology , radiation therapy , programming language , operating system
Prostate cancer patients who elect permanent implant brachytherapy at our institution are manually preplanned according to a set of guidelines that aim to achieve V 100 PTV > 98 % , 48 % < V 150 PTV < 55 % , and V 200 PTV < 18 % while focusing on complete coverage between the base and apical slices of the PTV and, if possible, creating a strong dose‐bias of the V 150 PTVto the posterior‐peripheral regions. A left‐right symmetric PTV is used, and this symmetry is reflected in planning by requiring a symmetrical needle distribution around the ‘D’ column, in which certain needle configurations are preferred. An inverse planning algorithm is under development to automate treatment planning that imitates the style of plans produced by manual planners trained at our institution. The current version of the algorithm uses a Bayesian network which is trained on a database of successfully implanted plans, generalizing our previous work which was based solely on needle placement frequency. The likelihood of a particular needle placement is modeled as a function of its spatial relationship to the PTV, as well as the distribution of neighboring needles. The latter conditional dependencies can be used to impose inter‐needle rules, such as forbidding adjacent needle placements. Effective needle sub‐configurations that appear frequently in the training set are learned by the network, which implicitly steers the optimization towards plans that resemble the manual ‘style’. The incorporation of this algorithm into clinical workflow will reduce workload, aid in training, improve consistency, and lay the foundations for automated planning in an intraoperative setting.