Click here to

Session: Machine Intelligence for Treatment Planning and Segmentation [Return to Session]

Cascaded Learning-Based Cone Beam CT Head-And-Neck Multi-Organ Segmentation

Y Lei, X Dai*, Z Tian, T Wang, J Zhou, J Roper, B Ghavidel, M McDonald, D Yu, J Bradley, T Liu, X Yang, Winship Cancer Institute of Emory University, Atlanta, GA


WE-G-BRC-2 (Wednesday, 7/13/2022) 2:45 PM - 3:45 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: As an on-board imaging system capable of capturing patients’ 3D anatomy at treatment position before delivery of each treatment fraction, cone beam CT (CBCT) has the great potential to guide online adaptive radiotherapy. This study aims to develop a novel deep learning-based model to automatically delineate organs-at-risk (OARs) to facilitate the CBCT-guided online adaptive process.

Methods: In the proposed cascaded learning-based model, a pretrained organ volume-of-interest (VOI) detection learning-based model, called fully convolutional one stage objection detector, is firstly used to extract informative features from CBCT and locate the VOIs of multiple OARs. The features outside the VOIs are excluded so that the cascaded learning-based model can ignore useless regions and reduce complexity. Then, a hierarchical block takes cropped feature maps as input to predict the enhanced feature maps aiming to reduce misclassification at mixed organ boundary. Finally, a segmentation network takes the enhanced feature maps as input and estimates the segmentation within detected VOIs. A retrospective study was conducted on a cohort of 30 head-and-neck cancer patients receiving proton therapy. Each patient includes multiple daily CBCTs acquired during treatment and a set of manual OAR contours are used as ground truth. Dice similarity coefficient (DSC) and mean surface distance (MSD) between predicted and manual contours were calculated to quantify the performance of the proposed method.

Results: Overall, the mean DSC values were 0.86±0.04, 0.72±0.19, 0.89±0.09, 0.85±0.13, 0.63±0.13, 0.89±0.08, 0.73±0.08, 0.75±0.11, 0.95±0.04, 0.79±0.09, 0.87±0.08, and 0.82±0.11 for brain stem, esophagus, larynx, mandible, optic chiasm, eye, lens, optic nerve, oral cavity, cochlea, parotid, and spinal cord, respectively. The mean MSD ranged from 0.55±0.14 mm to 1.24±0.98 mm for all OARs.

Conclusion: We have evaluated the effectiveness of our learning-based multi-OAR CBCT segmentation method, which offers a strategic solution for fast multi-OAR delineation, paving the way for CBCT-based online adaptive radiotherapy.


Not Applicable / None Entered.


IM- Cone Beam CT: Segmentation

Contact Email