Premium
SU‐C‐WAB‐05: Automatic Bladder Segmentation On CBCT for Plan Selection During Cervical ART
Author(s) -
van de Schoot A,
Schooneveldt G,
Wognum S,
Hoogeman M,
Chai X,
Stalpers L,
Rasch C,
Bel A
Publication year - 2013
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 - 10.1118/1.4813957
Subject(s) - segmentation , cone beam computed tomography , artificial intelligence , computer science , image registration , image segmentation , computer vision , volume (thermodynamics) , pattern recognition (psychology) , nuclear medicine , mathematics , medicine , radiology , computed tomography , image (mathematics) , physics , quantum mechanics
Purpose: Adaptive radiotherapy (ART) with a plan‐of‐the‐day (PotD) approach is suitable for cervical cancer patients, due to the highly mobile cervix‐uterus. PotD selection is not obvious given the poor soft‐tissue contrast of CBCT and bladder volume can be used as indirect measure of cervix‐uterus position. The aim of this study is to automatically segment the bladder on CBCT, in order to automate plan selection. Methods: Four cervical cancer patients, treated in prone position on a bellyboard, with planning‐CT and in total 25 CBCT (Elekta) images were included. A patient‐specific statistical training set was developed to guide automatic bladder segmentation. This training set was built from full and empty bladder contour interpolations and principal component analysis was applied to model deformation patterns. Bladder segmentation on CBCT was obtained by consistently deforming the planning contour using the training set while maximizing the cross‐correlation between directional gradient fields on both images. The segmentations could be improved by manually adding correction points. The training set could be expanded with the segmentation. The segmentations, acquired with expanded (ETs) or non‐expanded training set (NETs), were validated with manual delineations by volume comparison and contour‐to‐contour distance. The volume range between empty and full bladder was divided into three equally‐sized subranges, representing PotDs. The segmented and manually obtained volumes were classified into subranges, representing selected PotD. Results: Average volume correlations between segmentation and manual delineation were 0.84 (NETs) and 0.89 (ETs) and the mean absolute contour‐to‐contour distance was 0.59cm (NETs) and 1.21cm (ETs) using on average 2.48[0–4] (NETs) and 2.80[0–6] (ETs) correction points per segmentation. PotD selection using manual and segmented volumes (NETs and ETs) was equal in 92%. Conclusion: Bladder volumes of cervical cancer patients can be detected automatically on CBCT using a patient‐specific training set. The accuracy is sufficient for plan selection during cervical ART.