MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Classification Model Constrained 3D U-Net for Autosegmentation of Head and Neck Organs-At-Risk

S Zhang1*, H Wang2 , S Tian3 , X Zhang4 , J Li5 , R Lei6 , M Gao7 , C Liu8 , J Wang9 , R Yang10 , (1) Peking University Third Hospital, Beijing, ,(2) Peking University Third Hospital, Beijing, ,(3) Peking University Third Hospital, Beijing, ,(4) Peking University Third Hospital, Beijing, ,(5) Peking University Third Hospital, Beijing, ,(6) Peking University Third Hospital, Beijing, ,(7) LinkingMed, Beijing, ,(8) LinkingMed, Beijing, ,(9) Peking University Third Hospital, Beijing, ,(10) Peking University Third Hospital, Beijing,

Presentations

(Monday, 7/15/2019) 4:30 PM - 5:30 PM

Room: Exhibit Hall | Forum 2

Purpose: To develop a novel classification model constrained 3D U-Net model for autosegmentation of head and neck (HaN) organs-at-risk (OARs).

Methods: The autosegmentation was a two-step process, localization with the classification model using deep convolutional neural network and fine segmentation with 3D U-Net model. The classification model consisted of two parts, coarse classification model that was trained to classify CT slices into head region and neck region and fine classification model that was trained to divide the head region into four parts according to anatomical position of structures. Then 3D U-Net model focused on the corresponding region and fine segmented OARs. The CT image data of 170 head and neck patients were included in this study (150 for training and 20 for testing). OARs included brainstem, eyes, optic nerves, temporal lobes, mandible, temporomandibular joint (TMJ) and parotids. Dice similarity coefficient (DSC), 95% Hausdorff distance (HD) and segmentation time were evaluated.

Results: The average DSC and 95%HD values of all OARs were clinically acceptable, and better or comparable with literature reported results (brainstem: 0.88 and 2.27 mm, left eye: 0.89 and 1.43 mm, right eye: 0.87 and 1.59 mm, left optic nerve: 0.74 and 1.46 mm, right optic nerve: 0.72 and 1.57 mm, left temporal lobe: 0.82 and 3.55 mm, right temporal lobe: 0.80 and 3.79 mm, mandible: 0.89 and 1.80 mm, left TMJ: 0.77 and 1.64 mm, right TMJ: 0.74 and 1.64 mm, left parotid: 0.81 and 5.18 mm, right parotid: 0.84 and 3.91 mm for DSC and 95%HD, respectively). The average autosegmentation time was 6.77 seconds for all the OARs.

Conclusion: The developed classification model constrained 3D U-Net model demonstrated superior performance in both accuracy and efficiency. It had the potential for implementation into the clinical work with radiation oncologists approval.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by National Natural Science Foundation of China [81071237], Interdisciplinary Medicine Seed Fund of Peking University [BMU20160585] and Peking University Third Hospital Clinical Key Project [BYSY2018013].

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email