Click here to

Session: Multi-Disciplinary: Segmentation III [Return to Session]

Auto-Segmentation of Important Centers of Growth in the Pediatric Skeleton to Consider During Radiation Therapy Based On Deep Learning

W Qiu1*, W Zhang2, X Ma1, Q Zhou2, J Zhu1, (1) Department of Radiation Oncology Physics and Technology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Ji'nan, P.R. China, Jinan, Shangdong province, CN, (2) Manteia Medical Technologies, Milwaukee.


TH-IePD-TRACK 3-7 (Thursday, 7/29/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: Routinely delineating important skeleton growth centers is imperative to track and reduce exposure dose for pediatric neoplasms population with radiotherapy. The goal of this study is to explore network structure dimension effect on the efficiency and accuracy of deep learning segmentation models for skeleton growth plates. We devised a novel deep learning network called 2.5D UNet, that extends the semantics of adjacent slices to automatically segment the important centers of growing skeleton.

Methods: The craniofacial, shoulder and pelvis growth plates, with a total of 19 structures, were manually delineated by a radiation oncologist on axial, coronal, sagittal CT images from 75 child cancer patients. Based on the UNet network structure, three models of segmentation networks including 2D, 3D and proposed 2.5D are compared. 2.5D refers to the use of three consecutive image slices to predict the segmentation of the middle slice, which can increase the associated semantic information. Dice Similarity Coefficient (DSC) and Hausdorff Distance 95% (HD95) were used to evaluate the segmentation accuracy of each model, and time consumption was recorded respectively.

Results: Comprehensive evaluation of the 19 ROIs output by the three models of 2D, 3D and proposed 2.5D, the average DSC is 0.7972, 0.7975, 0.7987, and the average HD95 is 4.25, 1.88, 1.90, respectively. The HD95 performance of the latter two is clearly better than the former, while the DSC performance is not significantly different. The efficiency of the 3D model (23.5 hours of training time, 180 seconds of average inference time per patient) is significantly lower than 2D (3.1 hours, 45 seconds) and 2.5D (3.3 hours, 48 seconds).

Conclusion: The deep learning-based models to automatically delineate pediatric growth plates have a satisfactory segmentation performance. Considering segmentation accuracy and time efficiency, 2.5D is more appropriate for clinical application than 2D and 3D models.

Funding Support, Disclosures, and Conflict of Interest: the National Natural Science Foundation of China (grant numbers 81671785, 81530060 and 81874224), the National Key Research and Develop Program of China (grant number 2016YFC0105106), the Foundation of Taishan Scholars (No.tsqn201909140, ts20120505), the Academic promotion program of Shandong First Medical University (2020RC003, 2019LJ004), the Shandong Provincial Natural Science Foundation (ZR2016HQ09).



    Contour Extraction, Bone Structure, Radiation Protection


    IM/TH- image Segmentation: CT

    Contact Email