Semiautomatic bladder segmentation on CBCT using a population-based model for multiple-plan ART of bladder cancer

Abstract

The aim of this study is to develop a novel semiautomatic bladder segmentation approach for selecting the appropriate plan from the library of plans for a multiple-plan adaptive radiotherapy (ART) procedure. A population-based statistical bladder model was first built from a training data set (95 bladder contours from 8 patients). This model was then used as constraint to segment the bladder in an independent validation data set (233 CBCT scans from the remaining 22 patients). All 3D bladder contours were converted into parametric surface representations using spherical harmonic expansion. Principal component analysis (PCA) was applied in the spherical harmonic-based shape parameter space to calculate the major variation of bladder shapes. The number of dominating PCA modes was chosen such that 95% of the total shape variation of the training data set was described. The automatic segmentation started from the bladder contour of the planning CT of each patient, which was modified by changing the weight of each PCA mode. As a result, the segmentation contour was deformed consistently with the training set to best fit the bladder boundary in the localization CBCT image. A cost function was defined to measure the goodness of fit of the segmentation on the localization CBCT image. The segmentation was obtained by minimizing this cost function using a simplex optimizer. After automatic segmentation, a fast manual correction method was provided to correct those bladders (parts) that were poorly segmented. Volume- and distance-based metrics and the accuracy of plan selection from multiple plans were evaluated to quantify the performance of the automatic and semiautomatic segmentation methods. For the training data set, only seven PCA modes were needed to represent 95% of the bladder shape variation. The mean CI overlap and residual error (SD) of automatic bladder segmentation over all of the validation data were 70.5% and 0.39 cm, respectively. The agreement of plan selection between automatic bladder segmentation and manual delineation was 56.7%. The automatic segmentation and visual assessment took on average 7.8 and 9.7 s, respectively. In 53.4% of the cases, manual correction was performed after automatic segmentation. The manual correction improved the mean CI overlap, mean residual error and plan selection agreement to 77.7%, 0.30 cm and 80.7%, respectively. Manual correction required on average 8.4 markers and took on average 35.5 s. The statistical shape-based segmentation approach allows automatic segmentation of the bladder on CBCT with moderate accuracy. Limited user intervention can quickly and reliably improve the bladder contours. This segmentation method is suitable to select the appropriate plan for multiple-plan ART of bladder cancer.