Click here to

Session: Ultrasound, Optical, and Multi-modality Imaging [Return to Session]

Validation of a Commercial Artificial Intelligence Auto-Segmentation System for Head and Neck Treatment Site

P Tsai, S Huang*, R Press, A Shim, C Apinorasethkul, C Chen, L Hu, E Jang, H Lin, New York Proton Center, New York, NY

Presentations

MO-E115-IePD-F9-1 (Monday, 7/11/2022) 1:15 PM - 1:45 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 9

Purpose: Accurate delineation of organ-at-risks (OARs) represents the foundation of successful radiotherapy. The invention of artificial intelligence (AI) based auto-segmentation has been recently available with the intention of time-saving and improving accuracy. It is particularly critical for the complex treatment site, such as the head and neck. Even though various commercialized systems are available in auto-segmentation of medical images, validation on real-world clinical cases is needed. This study validated a commercial AI auto-segmentation system in MIM software (Version 7.1.5) and assessed its clinical application on head and neck (H&N) cases.

Methods: The Contour ProtégéAI in MIM was used for auto-segmentation creation with a total of 17 OARs for each case. The data was compared against the twenty-two head and neck benchmark datasets. The performance of the AI contours were evaluated using Dice Similarity Coefficient (DSC) and Hausdorff distance (HD) compared with the benchmark contours. In addition, the quality of each OARs was also visually evaluated by clinician.

Results: The MIM auto-segmentation system can achieve competitive accuracy compared to H&N benchmark dataset. The average DSC and HDmean were 0.71±0.07 and 1.8±2.0mm for all the OARs. 64.9% (243/273) of the OARs have DSC larger than 0.7, while 89.8% (226/273) of the OARs have HD less than 3mm. The majority of the OARs showed no significant difference with the benchmark contours, except for OARs of small volume or at the air-tissue interface, such as optic nerves, lens, chiasm, and larynx. Three cases show visible discrepancies on submandibular glands due to unique soft tissue morphology.

Conclusion: The Contour ProtégéAI can accurately delineate the OAR for auto-segmentation tasks on CT images with no user interventions. The platform can be implemented into a seamless workflow, enabling fast OARs segmentation in the head and neck region. Visual inspection of the contours is recommended for clinical use.

Keywords

Segmentation, CT, Radiation Therapy

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email

Share: