MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep Learning-Based Auto-Segmentation of OARs in Head and Neck CT Images

Z Shen1*, A Garsa1, S Sun2, N Bai2, C Zhang2, A Shiu1, E Chang1, W Yang1, (1) University of Southern California, Los Angeles, CA, (2) DeepVoxel Inc., Irvine, CA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Adaptive radiotherapy calls for fast and accurate delineation of organs at risk (OARs). However, manual OAR delineation in head and neck (HN) is time-consuming and user-dependent due to the complex anatomy and the large number of OARs involved. In this study, we aim to evaluate the clinical utility of a novel deep learning model, named attention-modulated U-Net (U?-Net), for HN radiotherapy planning.

Methods: The U?-Net was previously trained on a data set of 215 labeled CT images. U?-Net first detects regions containing OARs, and then upsamples image features only within the regions, instead of the whole volume as in the conventional U-Net. To implement the U?-Net in our institutional treatment planning system, we selected planning CTs from 20 HN patients previously treated with radiation. 16 OARs were auto-segmented, including brain stem, eyes, lenses, optic structures, spinal cord, mandible, oral cavity, parotids, submandibular glands, and esophagus. Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (95% HD) were used to evaluate the quality of OAR delineation. Clinically relevant dosimetric endpoints (max or mean dose) of the auto-segmented OARs were compared to those of the clinical contours.

Results: U?-Net was able to delineate all HN OARs within a few seconds, compared to over half an hour by human experts. U?-Net model achieved an average DSC of 72.73% (range: 35.00%-86.46%) and an average 95% HD of 6.69 mm (range: 2.86-13.43) across the 16 OARs, outperforming the traditional atlas-based auto-segmentation methods. Some dosimetric endpoints of the auto-segmented OARs were statistically different compared to the expert contours, but all met the clinical dose constraints.

Conclusion: U?-Net was able to auto-segment various OARs for HN radiotherapy planning in seconds and resulted in clinically acceptable dosimetry results. This auto-segmentation tool has the potential to improve the efficiency and consistency of OAR delineation.

Funding Support, Disclosures, and Conflict of Interest: SS, NB, and CZ are employees of DeepVoxel Inc.

Keywords

Segmentation, CT

Taxonomy

IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning

Contact Email