ESTRO 2024 - Abstract Book

S3748

Physics - Image acquisition and processing

ESTRO 2024

For 5 cervical cancer patients, 2 pelvic CT scans were used per patient that were acquired for radiation treatment planning purposes. One in which the patient had a full bladder, and one in which they had an empty bladder, acquired shortly after the full bladder scan. Both scans had in-slice resolutions ranging from 0.86mm to 1.08mm and a slice thickness of 3mm. On each scan with a full bladder, a reference set of 23 anatomically relevant landmarks was annotated by a radiation therapy technologist (RTT) and verified by a radiation oncologist, with 7 landmarks located on bony anatomy and 16 on soft tissue structures (see Figure 1). A group of 13 observers that consisted of 11 RTTs and 2 researchers were then given the protocol used for landmark placement. The given protocol instructs observers to indicate the corresponding location of each landmark as identified by the RTT on the full bladder scans, on the empty bladder scans as accurately as possible, and to record their confidence on a scale of three levels. Contours of the organs at risk were available for consultation. To quantify the inter-observer variability between different landmark placements for a patient’s empty bladder scan, we use their geometric median as an approximation of the true corresponding landmark location. We found that placements of landmarks on soft tissue for all patients deviate from the geometric median by a median 3D Euclidean distance of 3.0 mm (IQR [1.4-5.7]), and on bony anatomy by 1.4 mm (IQR [0.8-2.5]). Inter-observer variability of corresponding landmark placement on soft tissue was found to be significantly larger than that on bony anatomy (Mann-Whitney U test, p<0.001). Figure 2 shows the inter-observer variability of each anatomical landmark across patients. When the three dimensions are considered separately for the entire landmark set, landmarks deviated by a median of 0.7 mm in the left-right dimension, by 1.0 mm in the anterior-posterior dimension, and 1.1 mm in the cranial-caudal dimension. In general, inter-observer variability tended to increase (Spearman rank-order correlation coefficient 0.22, p<0.001) as the 3D Euclidean distance between the landmark's geometric median on one scan and their reference point on the other scan, after rigid registration, increases (i.e., as the magnitude of deformation in the region increases). We also found that an observer's confidence about a landmark placement is negatively correlated with the distance of the observer's placement from the geometric median of that landmark (Spearman rank-order correlation coefficient -0.42, p<0.001), indicating that participants often were able to correctly assess their own level of accuracy on an individual landmark placement. These findings provide important context to studies validating DIR methods, since such studies often rely on landmarks as a ground truth. Especially when validating DIR methods for clinical applications such as dose accumulation, voxel-level accuracy can be critical. However, the accuracy of a registration can only be measured to this level if the accuracy of the landmark placements themselves is known and sufficient. The inter-observer variability values reported in this study could serve as a lower bound for a meaningfully achievable level of accuracy, which facilitates the interpretation of DIR method validation outcomes. Results: