(Thursday, 8/2/2018) 7:30 AM - 9:30 AM
Room: Karl Dean Ballroom C
Purpose: It’s desirable to robustly validate the auto-segmented contours in online adaptive replanning (OLAR). We proposed a fast and automated patient-specific contour validation approach by using quantitative image texture features of the image of the day.
Methods: A total of 165 daily CTs collected using an in-room CT during CT-guided radiation therapy for 33 pancreatic cancer patients were utilized. The contours of pancreas head and duodenum generated using an auto-segmentation tool were tested with respective to the ground truth contours delineated by an experienced radiation oncologist. For each test contour, inward and outward shells were generated by eroding and expanding the contour by 3mm, separately. A total of 22 second-order Gray Level Co-occurrence Matrix(GLCM) based texture features were calculated for inner and outer shells, and then normalized to central region(20% in volume) in corresponding contours. We assume that, for an accurate contour, the texture values should be consistent for volumes enclosed by inward shell and drastically different in outward expansions. The appropriate texture features were identified as those with large different feature values between the in- and out-ward shells. The Receiver Operating Characteristic(ROC) curves were used to determine the thresholds for the chosen features to discern accurate and inaccurate contours.
Results: The combined Inner-ClusterProminence(InClusP), Inner-ClusterShade(InClusS) and Outer-ClusterProminence(OutClusP) features were identified to be appropriate to determine contour accuracy. To ensure high sensitivity(i.e,0.95), InClusP<1.1, InClusS<10 and OutClusP>20 were chosen for accurate pancreas head contours. Similarly, InClusP<7.1, InClusS<10 and OutClusP>20 were chosen for accurate duodenum contours. The corresponding specificity was 0.89 and 0.83 for pancreas
Conclusion: The proposed method of using quantitative texture features to evaluate and validate quality of contour can identify accurate or inaccurate contours with high sensitivity and specificity. The method can potentially replace the time-consuming manual process of checking contours during online adaptation, facilitating the routine practice of OLAR.
Keywords
Texture Analysis, Quality Control, Segmentation
Taxonomy
Contact Email
