Main Session
Sep 28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology

2463 - Feasibility of an AI-Assisted Radiotherapy Contour Quality Assurance Platform for Multicenter Cervical Cancer Clinical Trials in NRG Oncology

10:45am - 12:00pm ET
Poster Hall - Exhibit Hall A
Screen: 16
POSTER

Presenter(s)

Jyoti Mayadev, MD, FASTRO - UCSD, La JOlla, CA

H. Geng1, J. S. Mayadev2, H. Kim3, W. Deng4, D. Zamarin5, S. Ghamande6, C. R. Nwachukwu7, L. L. Holman8, C. Mathews9, D. O'malley10, A. B. Olawaiye11, E. Hopp12, R. O’Cearbhaill13, C. Aghajanian14, R. Schilder15, C. Leath16, and Y. Xiao17; 1University of Pennsylvania, Perlman School of Medicine, Philadelphia, PA, 2GOG Foundation, Philadelphia, PA, 3Department of Radiation Oncology, University of Pittsburgh Medical Center, Pittsburgh, PA, 4NRG Oncology, Buffalo, NY, 5Icahn School of Medicine Mount Sinai, New York City, NY, 6Augusta University, Augusta, GA, 7University of Texas Southwestern Department of Radiation Oncology, Dallas, TX, 8University of Oklahoma, Oklahoma, OK, 9Womens and Infant hospital, Providence, RI, 10Ohio State University, Columbus, OH, 11University of Pittsburgh, Department of Gynecologic Oncology, Pittsburgh, PA, 12Medical College of Wisconsin, Wisconsin, WI, 13MSKCC, New York City, NY, 14Memorial Sloan Kettering Cancer Center, New York, NY, 15Thomas Jefferson University, Philadelphia, PA, 16O'Neal Comprehensive Cancer Center, University of Alabama at Birmingham, Birmingham, AL, 17Department of Radiation Oncology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA

Purpose/Objective(s):

Variability in radiotherapy (RT) target and organ-at-risk (OAR) contouring remains a significant challenge in multicenter clinical trials, affecting protocol compliance, treatment quality, and trial outcomes. We assessed the feasibility and performance of an artificial intelligence (AI)–assisted RT contour quality assurance (QA) workflow in a multicenter trial for locally advanced cervical cancer (NRG-GY017), with the goal of informing scalable implementation in future clinical trials.

Materials/Methods:

Target volume definitions from the NRG-GY017 and standardized protocols were systematically compared. Discrepancies were resolved through consensus discussions with the principal investigator physician, resulting in harmonized definitions of subclinical target volumes. Expert-reviewed contours from 38 NRG-GY017 cases were used to train an AI model for segmentation of FDG-avid nodal gross tumor volume (GTVn).

A commercially available AI-based auto-segmentation platform was evaluated for delineation of female pelvic lymph node regions and OARs. A postprocessing workflow was developed to integrate AI-generated structures and generate clinical target volumes (CTVs) based on patient-specific disease extent and staging.

Quantitative contour metrics, including Dice similarity coefficient (DSC), Hausdorff distance, mean distance to agreement, and volumetric differences were calculated to compare AI contours with validated clinical contours.

An automated pipeline was implemented to generate standardized contour and dosimetric quality assurance review forms.

Results:

The contour comparison metrics plus/minus one standard deviations were used as thresholds (Table 1) for identifying contour deviations. When three out of four metrics for one contour falls beyond the threshold, a contour will be marked as deviations.

Conclusion:

Development of an AI-assisted RT contour quality assurance workflow for target and OARs was feasible. The integrated workflow demonstrated acceptable contour accuracy and improved efficiency when used in conjunction with principal investigator oversight. Expansion of the training dataset is expected to further enhance detection of contouring deviations and support scalable, standardized quality assurance in future NRG Oncology trials.

Kindly for Support & acknowledgment: Genentech, Member of the Roche Group
Table 1. Thresholds for contour deviation detection
Contour

 Dice

 HD (mm)

 MDA (mm)

 Volume Change(%)
CTV_4500

 0.7

 50

 10

 30%
GTVn

 0.48

 23

 5

 35%
Spc_Bowel

 0.3

 50.2

 14.3

 51%
Rectum

 0.71

 37.6

 5.4

 30%
Bladder

 0.87

 13.5

 1.3

 12%
Kidney_L

 0.85

 15

 2.6

 15%
Kidney_R

 0.85

 15

 2.6

 15%
SpinalCord

 0.96

 35

 1

 15%
Duodenum

 0.7

 40.6

 6.5

 30%
Liver

 0.82

 35

 3.5

 15%
Femur_Base_L

 0.75

 37.9

 5.4

 12%
Femur_Base_R

 0.75

 37.9

 5.4

 14%