(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: To establish a knowledge-base of normal tissue structure characteristics for use in automated structure evaluation.
Methods: Geometric characteristics of 4436 normal tissue structures were gathered from patients that were treated clinically with photon IMRT or VMAT plans. Gathered metrics include the structure volume, mean area per slice, the extent of the structure in the three cardinal directions, the coordinate of the geometric center, and the maximum dose to the structure from the delivered treatment plan. Characteristics of non-anatomical structures (those used for optimization, avoidance, etc.) were not recorded. From the collected data, a structure knowledge-base was created by determining the mean and standard deviation of each geometric characteristic. Data were separated by structure type and patient gender. An additional 800 contours were manually reviewed by a physicist and categorized as acceptable or unacceptable. For those marked unacceptable, the contouring error category was also recorded. A receiver-operating characteristic (ROC) curve for detecting contouring errors was generated by evaluating the manually-reviewed dataset metric q with the knowledge-base (KB) values using a variable decision threshold m.
Results: Of the 800 manually-reviewed contours, 93 were marked unacceptable (under-contoured=51%, over-contoured=39%, incorrect structure=8%, uninterpolated=3%). Structure volume outperformed mean area and superior/inferior extent in detecting contouring errors (AUC=95.0%, 81.4%, and 73.3% respectively). Using a decision threshold of m=0.8, structure volume was able to detect contouring errors with sensitivity=87%, specificity=94%, and accuracy=93.4%.
Conclusion: A knowledge-base of normal tissue geometric characteristics has been established. This knowledge-base was then used to demonstrate that geometric characteristics of normal tissue structures, particularly the structure volume, can be used to detect contouring errors with excellent accuracy and specificity.
Funding Support, Disclosures, and Conflict of Interest: MCS reports, Varian
Keywords
Segmentation, ROC Analysis, Statistical Analysis
Taxonomy
Contact Email
