ESTRO 36 Abstract Book

S491 ESTRO 36 _______________________________________________________________________________________________

adjacent contours, 3) conformity index (CI) between adjacent contours, 4) presence of air or bone across the line of the contour, 5) presence of air or bone within 5 mm outside of the contour boundary, and 6) presence of spacing > 20 mm between adjacent contour points. The threshold values for the metrics 1-3 were calculated from the rectum contours drawn by oncology experts on 315 pelvic kVCT scans, where we used 6 mm superior- inferior contour spacing to match the slice spacing of the IG scans. The settings for metrics 4-6 were determined empirically. Our software developed in Python 2.7 analysed the DICOM RTSTRUCT and IG scan data, applied the metrics and recorded the evaluation results in a spreadsheet. A contour was marked as “error” if any of the thresholds defined in the metrics was triggered. Results

Conclusion There was low interobserver variability across all centres. Low rates of protocol deviations ensured high compliance of the participating centres. Targets were adequately and homogeneously covered in the majority of patients. Dose parameters were comparable between OD and GS and confirmed that interobserver variability did not influence treatment outcomes.

The automatic evaluation of 11519 contours for 33 patients took 6 minutes on a computer with 8 GB RAM and 1.6 GHz Intel Xeon CPU. The evaluation results were compared to the errors recorded by a human observer, and confusion matrices were calculated. The mean error prevalence in the observer evaluation was 0.29 ± 0.1. Our algorithm achieved a mean sensitivity of 0.84 ± 0.1 (range [0.58 – 1.0]) and a mean specificity of 0.88 ± 0.1 (range [0.51 – 1.0]). One patient data set totalling 339 slices was evaluated with a sensitivity and specificity of 1.0. Conclusion Metric-based evaluation of rectum contours is a feasible alternative to evaluation of contours by a human observer. It provides an unbiased contour classification and detects over 80% of typical errors in the contours. The method can be used to assess the performance of automated contouring tools and to aid the development of improved contouring software. PO-0893 Improving CBCT image quality for daily image guidance of patients with head/neck and prostate cancer I. Chetty 1 , P. Paysan 2 , F. Siddiqui 1 , M. Weihua 1 , M. Brehm 2 , P. Messmer 2 , A. Maslowski 3 , A. Wang 3 , D. Seghers 2 , P. Munro 2 1 Henry Ford Health System, Radiation Oncology, Detroit, USA 2 Varian Medical Systems Imaging Laboratory GmbH, Image Enhancement and Reconstruction, Baden-Daettwil, Switzerland 3 Varian Medical Systems- Inc., Oncology Systems, Palo Alto, USA Purpose or Objective Image quality of on-board CBCT imaging in radiation therapy generally falls short of diagnostic CT in particular

Poster: Physics track: Images and analyses

PO-0892 Automatic quality assurance of rectal contours on image guidance scans M. Romanchikova 1 , D.I. Johnston 1 , M.P.F. Sutcliffe 2 , K. Harrison 3 , S.J. Thomas 1 , J.E. Scaife 4 , N.G. Burnet 4 1 Cambridge University Hospitals, Medical Physics and Clinical Engineering, Cambridge, United Kingdom 2 University of Cambridge, Engineering, Cambridge, United Kingdom 3 University of Cambridge, Physics, Cambridge, United Kingdom 4 University of Cambridge, Oncology, Cambridge, United Kingdom Purpose or Objective Assessment of the quality of contours produced by automatic methods is labour-intensive and inherently dependant on the skills of the evaluator. The utilisation of these contours in radiotherapy requires objective quality metrics and efficient tools for contour quality assurance. We present a method to determine the quality of automated rectum contours on daily image guidance scans (IG). Material and Methods We analysed 11519 automatically produced rectum contours on 1062 pelvic IG scans of 33 prostate cancer patients. Each contour was evaluated by 1) a trained clinician and 2) an automated classification software that applied a set of binary and numeric metrics to each contour. The metrics included 1) centre-to-centre contour distances, 2) differences in contour areas between