Robust and reliable stripe detection for CBCT

TitleRobust and reliable stripe detection for CBCT
Publication TypeConference Paper
Year of Publication2009
AuthorsArns, A, Blessing, M, Stsepankou, D, Wertz, H, Lohr, F, Hesser, J, Wenz, F
Conference NameDEGRO

Purpose: Combined kV-MV cone-beam CT reduces scanning time by up to 75% compared to the standard scanning procedure. Hence, this technology is enabling on-board imaging for IGRT of lung cancer within a breathhold. 60 kV and 40 MV projections are acquired. MV projections are mapped onto kV energy by histogram adaptation and then reconstructed by filtered back projection. Since readout and irradiation are not synchronized, brightness variations on the projection data manifest as stripes. Robust and problem independent stripe detection is thus a necessary processing task for solving this problem and is discussed in detail. Methods and Materials: Projection data was acquired from an Elekta Synergy 6MV Linac. In the processing workflow, the variance of gradient locations along the stripe direction was used to generate a profile which was then selected for locating stripes. Although these profiles showed articulated peaks in the presence of stripes, their amplitude and width were not uniform indicating that thresholding was neither sufficient, nor stable for extracting them from the background. Standard methods for automatic threshold detection thus failed. In our approach, the variance of high gradient locations coincided with peak maxima but other structures in the image led to maxima as well, although being smaller. Since the stripes had a certain width, threshold selection failed. Non-maximum-suppression overcame this problem leading to a large gap between peaks in the profile originating from stripes and those coming from other structures. Finding the largest gap separating these two classes was found to be a robust and reliable method for classifying stripes and other structures. Results: Using this automated processing, we determined stripes for four different projection data sets with each 150-160 images. The sensitivity of this method was 92.8% while the specificity was 100%. Running on a PC Pentium® 4 CPU 3GHz and non-optimized Matlab, processing time per 754x754 pixel image was 3.9 ± 0.9 [s]. In case of very bright background structures, not all stripes were detected, because the magnitude of the stripes only marginally differed from the magnitude of these other structures. Therefore, the variance profile was very unsteady regarding to peak maxima and thus the threshold was set too high. Conclusions: Using this automated stripe detection, the kV-MV-preprocessing and reconstruction software now automatically reconstructs acquired projections into volumes. High contrast structures like tumors in the lung can then be easily detected in the reconstructed volume and thus potential misalignment of patients can be found.

Citation Key