162 - Pilot Clinical Implementation and Evaluation of an AI-Based Decision Support Tool for CBCT-Guided Radiotherapy Image Review

Purpose/Objective(s): Anomalies in CBCT-guided radiotherapy, such as setup misalignments and soft-tissue variations, can signal treatment deviations potentially impacting quality and safety. We report the first prospective clinical application of an unsupervised anomaly detection framework (ADF) designed to automatically flag treatment fractions requiring in-depth review.

Materials/Methods: The ADF, using a variational autoencoder (VAE)-based CBCT inpainting technique, was previously developed and validated using registered simulation CT (simCT) and setup CBCT datasets from 1130 patients. The ADF generates an anomaly score derived from image similarity metrics between the simCT, daily CBCT and VAE-generated CBCT images, expressed as a percentile of the global distribution of scores. Initial validation was performed retrospectively on an unseen dataset of 243 patients, including 7 known setup incidents and simulated setup errors (one 2 cm translational error per patient). For prospective evaluation, daily setup CBCTs were processed automatically by the ADF nightly, and expert physicists reviewed the fractions with anomaly scores above the 95^th percentile the next day. Cases judged to substantially deviate from the original treatment goals were escalated to the attending physician. Flagging frequency, case escalations, plan revisions, and missed events were recorded.

Results: During the initial validation phase on known and simulated errors, the ADF achieved a specificity of 74% for a fixed sensitivity of 95%. During prospective evaluation, the ADF analyzed a total of 2881 treatment fractions (341 patients) between November 2025 and February 2026. 182 fractions (6%) from 67 unique plans (61 patients) were flagged for expert review. No gross misalignment was observed during this timeframe. Three cases demonstrated clinically significant anatomical deviations and were escalated to the physician; treatment adjustment was triggered in two cases (due to tumor growth and ascites, respectively). The third escalated case occurred near treatment completion and did not prompt a revision. Other flagged cases often reflected subpar image quality or variations in the structures adjacent to the target area (e.g. visceral organs). During the same period, three unflagged plans were adapted due to tumor growth (40^th, 72^nd and 94^th percentile respectively).

Conclusion: These findings indicate that the ADF has the potential to support routine physicist and physician image reviews by highlighting atypical cases, thereby enhancing treatment oversight and quality.