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

2642 - AI-Empowered Unsupervised Radiotherapy Workflow for Rectal Cancer

10:45am - 12:00pm ET
Poster Hall - Exhibit Hall A
Screen: 26
POSTER

Presenter(s)

Luqi Wang, MD Headshot
Luqi Wang, MD - Peking University Third Hospital, Beijing, Beijing

L. Wang1,2, R. Peng1,2, X. Li1,2, Y. Pan1,2, M. Wang1,2, M. Yang3, W. Zhang3, L. Jia3, and H. Wang1,2; 1Department of Radiation Oncology, Peking University Third Hospital, Beijing, China, 2Beijing Key Laboratory for Interdisciplinary Research in Gastrointestinal Oncology (BLGO), Peking University Third Hospital, Beijing, China, 3Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China

Purpose/Objective(s):

To develop a baseline risk model, using information available at initial simulation, to predict whether a rectal cancer patient can complete an unsupervised CT-linac workflow integrating CT simulation, auto-segmentation, auto-planning, and treatment delivery.

Materials/Methods:

Two radiation oncologists and two medical physicists independently rated auto-plan acceptability on a four-level scale (excellent/good/fair/poor). Eligibility for unsupervised delivery was defined as no “poor” ratings and =3/4 “excellent” or “good” ratings. Candidate predictors included baseline clinical variables and simulation-derived imaging features.

Base on the retrospective grading and clinical data, a machine learning model was built, including xgboost, random forest and logistic methods, for the purpose of assisting doctors in determining whether to perform the unsupervised AIO treatment. The retrospective data were split into training and testing cohort (4:1), and 18 features of each patient were filtered into the 10 based on descending correlation coefficient. The accuracy of each method was compared.

Results:

Between May 2024 and November 2025, 80 patients were accrued; 26/80 (32.5%) met unsupervision eligibility in CT-linac workflow. Correlation screening identified the top 10 features associated with eligibility as EMVI (r=0.260), T stage (r=-0.213), N stage (r=-0.160), MRI-measured tumor-to-anal margin distance (cm; r=-0.156), total tumor length (r=0.144), pelvic small-bowel volume (r=-0.143), bladder volume (r=0.135), overall stage (r=-0.109), MRF (r=0.085), and prior pelvic surgery (r=-0.079). Using these predictors, the classification accuracy was 0.750 for gradient-boosted trees, 0.750 for logistic regression, and 0.625 for random forest, supporting the feasibility of machine learning-based decision support for unsupervised eligibility triage in the AIO workflow. As shown in Table 1, automation performance was robust, with high auto-segmentation agreement and a higher plan-quality achievement rate in the eligible group.

Conclusion:

Using baseline clinical and imaging-derived features, the model aims to identify patients who are more likely to be eligible for unsupervised radiotherapy (e.g., those with EMVI, lower T/N stage, a shorter MRI-measured tumor-to-anal margin distance, and a longer total tumor length), thereby reducing patient waiting time and clinician workload.
Table 1 Automation performance by unsupervised eligibility

Notes: DSC——Dice similarity coefficient. CTV45 and CTV50 were the clinical target volumes prescribed to 45 Gy and 50 Gy, respectively. Achievement rate (%) denotes the percentage of cases meeting all 16 prespecified plan-quality criteria.

Metric

 eligibility (n=26)

 ineligibility (n=54)

 P value
Auto-segmentation

DSC-GTV

 0.897 ± 0.077

 0.806 ± 0.169

 p=0.017
DSC-CTV45

 0.954 ± 0.021

 0.932 ± 0.036

 p=0.008
DSC-CTV50

 0.923 ± 0.042

 0.871 ± 0.069

 p < 0.001
Auto-planning

Achievement rate (%)

 69.7 ± 9.0

 57.2 ± 14.1

 p < 0.001