Main Session
Sep 29
PQA 05 - Breast Cancer, International/Global Oncology

2938 - Feasibility Study of AI-Assisted Radiation Therapy Contour QA for a Breast Cancer Multi-Center Clinical Trial

03:00pm - 04:00pm PT
Hall F
Screen: 15
POSTER

Presenter(s)

Huaizhi Geng, PhD - University of Pennsylvania, Philadelphia, PA

H. Geng1, J. R. White2, E. E. R. Harris3, J. G. Bazan Jr4, R. S. Cecchini5, G. P. Chen6, J. Betler7, S. A. Seaward8, H. J. Sharp9, M. A. Proper10, K. Sarma11, N. Le-Lindqwister12, A. Tinger13, T. Biswas14, J. Lyons15, B. M. Anderson16, W. Sikov, MD17, E. P. Mamounas18, and Y. Xiao19; 1University of Pennsylvania, Perlman School of Medicine, Philadelphia, PA, 2Department of Radiation Oncology, University of Kansas Medical Center, Kansas City, KS, 3St. Luke's University Health Network, Easton, PA, 4Department of Radiation Oncology, City of Hope National Medical Center, Duarte, CA, 5NRG Oncology Statistical and Data Management Center; Department of Biostatistics and Health Data Science, University of Pittsburgh, Pittsburgh, PA, 6Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI, 7Allegheny Health Network Cancer Institute, Pittsburgh, PA, 8Kaiser Permanente Oncology Clinical Trials, Vallejo, CA, 9Atrium Health Levine Cancer Institute, Wake Forest University School of Medicine, Charlotte, NC, 10Montana Cancer Consortium, Billings, MT, 11Department of Radiation Oncology, Carle Foundation Hospital, Urbana, IL, 12Illinois CancerCare (Heartland NCORP), Peoria, IL, 13Bassett Healthcare Network, Cooperstown, NY, 14UF Health Cancer Center, University of Florida, Gainesville, FL, 1515University Hospitals Seidman Cancer Center / Case Western Reserve University / University Hospitals of Cleveland, Cleveland, OH, 16University of Wisconson School of Medicine and Public Health, Madison, WI, 17Women and Infants Hospital of Rhode Island; Warren Alpert Medical School of Brown University, Providence, RI, 18AdventHealth Cancer Institute, Orlando, FL, 19Department of Radiation Oncology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA

Purpose/Objective(s): This study evaluates the integration of Artificial Intelligence (AI) in the quality assurance (QA) process of radiotherapy contouring, specifically for radiotherapy data submitted to the multi-center clinical trial.

Materials/Methods: Contours from 54 cases were initially reviewed independently by three physicians, who assigned a review score to each contour (Score 1: per protocol; Score 2: acceptable variation; Score 3: unacceptable deviation). A commercial AI-driven auto-segmentation platform was evaluated, and a workflow was developed for post-processing and comparing AI-generated segmentations with the submitted data. Preliminary thresholds for contour comparison indices (Dice, Hausdorff distance, mean distance to agreement, volume differences) were established to identify potential deviations. Following physician meetings and feedback, clinically significant contour deviations were identified, including overlapping of clinical/planning target volumes (CTVs/PTVs) with the chest wall, skin, ribs, lungs and heart, as well as over-contoured or under-contoured breast tissues. Corresponding quantitative thresholds for detecting these deviations were defined and incorporated into the AI workflow. Subsequently, the AI-based contour reviews were compared with the PI reviews.

Results: Additional quantitative thresholds were incorporated into the workflow, including:

  • Overlapping breast tissue with the patient’s chest wall (10% of the breast volume);
  • Overlapping of PTV_WB_EVA with ribs (5cc)
  • Failure of CTV_WB/PTV_Lump to encompass the entire lumpectomy bed
  • Failure of breast organ at risk (OARs) to include the entire PTV_Lump_EVA
Integrating these quantified criteria enhanced the AI tool’s sensitivity in detecting Score 3 contour deviations. Sensitivity improved from 66.7% to 80% for CTV_WB, from 20% to 80% for PTV_Lump, from 40% to 80% for PTV_WB, and from 50% to 100% for breast tissue as an OAR. Additionally, the AI review identified two CTV_WB as well as five breast OAR contour violations that were initially missed by the clinicians.

Conclusion: Preliminary results demonstrate that AI's effectiveness in identifying suboptimal contours through the application of quantitative criteria, plus incorporating PI feedback, has the potential to increase accuracy and consistency. This study also indicates that AI-based contour QA may have the potential to reduce inter-rater variability and enhance the efficiency of trial QA or clinical contour evaluation. Clinical Trial number NRG-BR007.