Automatic breast-surface segmentation of deep inspiration breathold Cone-beam CT scans

Purpose: Breast tissue is highly rjeformahle and after breast conserving surgery considerable inter-fraction surface deformations occur during a course of radiotherapy. Depending on the magnitude of these shape changes these patients might need to be re-planned to obtain more accurate dose delivery. The deformations can be visualized with curvature maps on the external breast surface mesh. Although automatic contouring tools are widely available for CT, this is not the case for Cone-beam CT (CBCT). where Hounsfield unit accuracy is compromised by scatter and beam hardening which hamper the segmentation process. The purpose of this study was the development of an automated segmentation tool that works for CBCT data. Materials: To extract the external breast surface we first crop the conical caps, effectively removing isolines that connect the outer surface with inner structures like the lungs. According to the grayscale intensity histogram of the scan data, a segmentation level was automatically chosen to be the average of the air- and tissue peak in the scan histogram. Based on this level we applied the marching cubes algorithm, resulting in a 3D surface mesh. The 1ooi was evaluated on CBCT scans of five left-sided breast-cancer patients that were acquired during a deep inspiration breath-hold for setup verification. As a reference, the external surfaces were manually delineated by a RT technician. The distance from the automatic segmented mesh to the manually delineated mesh was calculated and the overall accuracy of the segmentation was quantified in terms of root-mean-square (RMS). Results: The RMS over the patients ranged from 0.15 cm to 0.21 cm with an average of 0.19 cm. It should be noted that there was only one observer and no computable variability in the delineations, therefore these results are an upper limit of the segmentation accuracy. Automatic segmentation takes about 8 sec on an Intel Core 2 Duo 3.16GHz PC, which is fast enough to be used in clinical practice. Conclusions: We developed a tool that automatically segments the external breast surface from CBCT data. The segmentations are accurate within 0.2 cm and useful for current and ongoing breast-cancer research. Automatic segmentation for free breathing breast cancer patients and application of segmented surfaces for biomechanical models are subject of further study.