Click here to

Session: CT and CBCT: New Technologies, Algorithms, and Emerging Applications II [Return to Session]

Leveraging the Elliptical Shape of the Uterocervix On Semi-Axial Cross-Sections for Improved Deep-Learning Segmentation On Cone-Beam CT

S Mason1*, L Wang2, K Zormpas-Petridis2, M Blackledge2, S Lalondrelle1, H Mcnair1, E Harris2, (1) Royal Marsden NHS Foundation Trust, Sutton, SRY, GB, (2) Institute Of Cancer Research


SU-F-201-3 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room 201

Purpose: Automated uterocervix segmentation on CBCT could assist with plan-of-the-day selection and online replanning for cervical cancer radiotherapy but remains a challenging task. For the first time, we test the hypothesis that semi-axial planes that represent the uterocervix as a series of regular elliptical cross-sections may be leveraged to improve the accuracy of uterocervix contouring on CBCT using deep-learning with a U-Net architecture rather than using conventional axial planes.

Methods: Patients: CBCT scans from 10 cervical cancer patients during radical radiotherapy. Gold Standard: An expert manually contoured the uterocervix on 4-10 CBCTs per patient. The CBCTs from 7, 2, and 1 patient(s) formed the training (1,628 images), validation (204 images), and test (204 images) sets respectively. Input: (a) The semi-axial planes for each CBCT were determined using an initialization step; the uterocervix was manually contoured on a central sagittal slice and annotated with four points to split the contour into quadrants. This inherently provides upper/lower bounds for the elliptical contour on each semi-axial plane and this information was included as an additional channel. (b) For the axial inputs, evenly spaced slices spanning the length of the uterus were selected. Augmentation: Random left-right flipping and left-right translation by up to 20 pixels quadrupled the training and validation sets.

Results: The mean +/- standard-deviation of the Dice Similarity Coefficient between the manual and U-Net segmentations was significantly higher for the semi-axial model (0.84 +/- 0.06) compared with the axial model (0.54 +/- 0.10) (p << 0.001) on the test set.

Conclusion: We have simplified the segmentation of a complicated 3D organ whose axial cross-sections are irregularly shaped and subject to splitting to segmenting 2D elliptical cross-sections. A highly significant improvement in segmentation accuracy was found (reaching that of inter-observer contour agreement on CBCT), even with relatively little training data.

Funding Support, Disclosures, and Conflict of Interest: Funding: (1) National Institute for Health Research and (2) Health Education England and NIHR Biomedical Research Centre at The Royal Marsden NHS Foundation Trust and the Institute of Cancer Research. Conflict of Interest: We are recipients of financial support from Elekta Ltd for collaborative work related to this project.


Cone-beam CT, Segmentation, Radiation Therapy


IM- Cone Beam CT: Segmentation

Contact Email