Fast and Robust Automated Segmentation of the Cervix-Uterus Structure in CT-Images Driven by Patient-Specific Motion-Models
L Bondar*, B Heijmen, M Hoogeman, Erasmus MC Daniel den Hoed Rotterdam, The NetherlandsMO-G-BRA-4 Monday 5:15:00 PM - 6:00:00 PM Room: Ballroom A
Purpose: The aim was to develop and test a novel automated intra-patient segmentation method of the cervix-uterus structure in CT-scans of cervical cancer patients to facilitate the online selection of the best plan-of-the-day based on in-room acquired CT-scans. Current automated segmentation methods for pelvic organs, which use statistical shape models and require a large training set, are unsuitable for cervix-uterus due to large inter-patient variability in shape and position.
Methods: An automated segmentation method was implemented that adapts a closely initialized surface of the cervix-uterus to boundaries in the new image. The novel idea was to use patient-specific motion-models derived from only two pretreatment CT-scans to initialize and drive the segmentation process. The cervix-uterus surface was initialized by using a 3D patient-specific cervix-uterus model that predicts the shape and position of cervix-uterus based on bladder volume, a 1D model predicting the bladder volume based on a manually marked bladder top, and implanted markers. The segmentation method was tested on 13 patients that had 9-10 CT-scans acquired at pretreatment and after 40 Gy. For each patient, two pretreatment CT-scans (full and empty bladder) were used for model construction and others were used for testing. The overlap between manually delineated and automatically segmented cervix-uterus structures was quantified by the Dice coefficient.
Results: Marking the bladder top and markers required minimal user intervention (<1 min). The overlap between the manually delineated and the initialized cervix-uterus structures was 82±7% for pretreatment and 71±11% for after 40 Gy data. The automatic adaptation of the initialized structure to image boundaries increased the overlap to 87±3% for pretreatment and 80±13% for after 40 Gy. The automatic segmentation method required 2±0.5 min.
Conclusions: A fast and robust automated segmentation method was developed that could support plan selection in online adaptive radiotherapy for cervical cancer patients.