G Pednekar¹, J Udupa², S Nag¹, Y Tong², T Liu², J Lukens³, A Berman³, D Mihailidis³*, J Stambaugh³, C Simone⁴, I Choi⁴, R Press⁴, C Robinson⁵, W Thorstad⁵, S Jabbour⁶, S Kim⁶, M Reyhan⁶, J Camaratta¹, S Owens¹, D Torigian², (1) Quantaras, Philadelphia, PA, (2) Medical Image Processing Group, Department of Radiology, University of Pennsylvania, Philadelphia, PA, (3) Department of Radiation Oncology, University of Pennsylvania, Philadelphia, PA, (4) New York Proton Center, New York, NY, (5) Washington University School of Medicine, Department of Radiation Oncology, Saint Louis, (6) Department of Radiation Oncology, Rutgers Cancer Institute Of New Jersey, Rutgers University, New Brunswick, NJ

• 
  D Mihailidis

PO-GePV-M-205 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Consensus-guidelines exist to guide practitioners on how best to perform contouring of organs-at-risk (OARs) on medical images. Yet, manually-assisted contouring of OARs is prone to variability, inaccuracy and is labor-intensive, especially for cases with image-distortions e.g., low-contrast, artifacts, pathology, surgery, and radiation treatment. We hypothesize: OAR auto-contours generated by novel Automatic-Anatomy-Recognition(AAR), which combines information from existing consensus-guidelines with artificial intelligence, is non-inferior in accuracy to those created in standard clinical practice and significantly saves human contouring time for challenging cases.

Methods: In a prospective multicenter clinical-evaluation at 4-US academic centers, contours for 16-head & neck(HN) and 10-thoracic(TH) OARs were auto-generated by AAR using CT-scans from 214-adult cancer patients undergoing RT-planning. Expertly drawn ground-truth OAR contours, based on existing contouring-guidelines, were created as independent reference standard. Standard-of-care clinical OAR contours were created by dosimetrists/ radiation oncologists and contouring times were recorded. Adjustments of AAR auto-contours were also performed as needed for RT-planning and recorded. Contouring time saved using AAR vs. clinical contouring was calculated for standard and challenging cases. Dice-coefficients for automatic and clinical contours were calculated and compared using non-inferiority test.

Results: Accuracy: For 11-OARs(6-HN, 5-TH), AAR-contours were equivalent to clinical-contours, for 8-OARs(6-HN, 2-TH), AAR-contours accurately represented the consensus guidelines. For 7-OARs(4-HN, 3-TH), AAR-contours were significantly better than clinical-contours denoting most-challenging OARs for human contouring using current clinical systems. AAR reduced clinical contouring time by (a) For standard cases: TH: 65%(27.30minutes) and HN: 55%(18.71minutes); (b) for challenging cases: TH: 74%(39.62minutes) and HN: 61%(29.71minutes).

Conclusion: Accuracy of AAR auto-contouring of HN and TH OARs is non-inferior to clinical contouring suggesting utility in promoting conformity and uniformity of OAR contouring relative to existing consensus guidelines. OAR contouring for cases with image distortions is more prone to human variability, inaccuracy, time-inefficiency thereby denoting significant improvement in clinical efficiency for those cases.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email