MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

AI-Seg: An Artificial Intelligence (AI)-Based Automatic Organs at Risk(OAR) Contouring Platform for Head and Neck Cancer (H&N) Radiotherapy

J Wu*, P Lynch , J Shah , W Lu , X Gu , UT Southwestern Medical Center, Dallas, TX

Presentations

(Sunday, 7/14/2019) 2:00 PM - 3:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: Modern head and neck (H&N) cancer radiotherapy is typically performed using intensity modulation therapy, which demands accurate organs at risk (OARs) delineation to ensure treatment quality. Manual contouring, as a standard clinical practice, is tedious, error-prone, suffering from inter- and intra- observer variations. This study aims to build an AI-based automatic OARs segmentation platform (AI-Seg) and integrate it with Eclipse treatment planning system.

Methods: The core of AI-Seg is our home-developed Enhanced 3D UNet (EnUNet) model. EnUNet recursively segment 26 H&N OARs in three levels. OARs with larger volumes and high contrast to their surroundings are assigned to level I and segmented first. The subsequent level II and III OARs are segmented with the constraints from prior levels segmentation results. The integration to Eclipse is achieved with a file-system-based client-server infrastructure. The AI-Seg computer server and Eclipse share a data folder. The computer server monitors Eclipse request and initiate auto-segmentation as soon as it receive a valid request from Eclipse. The final segmentation results are saved in the shared folder.

Results: The AI-Seg were successfully constructed and integrated with Eclipse TPS, where Eclipse can send request and segmentation results from AI-Seg can be visualized in the Eclipse. EnUNet were trained and tested on a database with 180 H&N patients. So far, we have 13 OARs segmented. Using manual delineation as gold standard, we achieved mean Dice Similarity Coefficient (DSC) are 76.2%(Eye_R), 80.3%(Eye_L), 79.8%(BrainStem), 84.3%(Cerebellum_R), 82.2%(Cerebellum_L), 81.4%(SpinalCord), 90.1%(Mandible), 73.3%(Parotid_R), 76.7%(Parotid_L), 77.2%(Submandibular_R), 80.2%(Submandibular_L), 86.1(Masseter_R), 87.3%(Masseter_L), respectively.

Conclusion: We have built AI-Seg platform and integrate it with Eclipse TPS. We will complete and further improve AI-Seg platform performance and apply it for clinical use.

Funding Support, Disclosures, and Conflict of Interest: the seed grant of Radiation Oncology Department at University of Texas Southwestern Medical Center.

Keywords

3D, Segmentation, Rotational Therapy

Taxonomy

IM- Cone Beam CT: Segmentation

Contact Email