1130 - Real-Time AI Compensation in Surface Guided Radiotherapy: A Multicenter Randomized Trial for Submillimeter Setup Accuracy and DIBH Adherence
Presenter(s)
F. Zhang1, Z. Liu2, F. Bai3, J. Li4, L. Xu5, and L. Zhao6; 1State Key Laboratory of Holistic Integrative Management of Gastrointestinal Cancers and Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xi’an, China, China, 2Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xi’an, China, 3Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, xi'an, Shaanxi, China, 4Department of Radiation Oncology,Xijing Hospital,Air Force Medical University CN, Xi’an, China, 5Department of Radiation Oncology, Xijing Hospital,Air Force Medical University( Fourth Military Medical University), Xi'an, China, 6Department of Radiation Oncology, Xijing Hospital, Air Force Medical University, xi an, shan xi, China
Purpose/Objective(s): Current surface-guided radiotherapy (SGRT) faces dual challenges: 1) Suboptimal intrafraction motion compensation (baseline error >2mm in 23% cases) with conventional laser alignment systems 2) Significant variability in DIBH compliance rates (55-78% success). We innovatively integrated LSTM temporal prediction with GAN-based respiratory coaching to address these limitations. This study hypothesized that the AI-SGRT system integrating LSTM timing prediction with GAN urinary attraction can significantly reduce radiotherapy positioning error (MAE < 0.2 mm) and improve DIBH adherence (? = 20%), aiming to verify its prospective RCT Real-time correction efficacy of multi-site tumors, optimization effect of patient respiratory training and cardiac dose protection advantages.
Materials/Methods: This multicenter stratified RCT (KY20232161-F-1) enrolled 150 patients (73% female, median age 61, KPS=80) between 1/2024-12/2024. Stratification included: - Breast cancer (n=60): T1-2N0M0, 65% left-sided - H&N cancer (n=50): 82% HPV+ oropharynx, 68% stage III-IV NPC - Prostate cancer (n=40): Gleason 7-9, PSA 4-20ng/mL -Intervention: 1) LSTM Motion Compensator trained on 100k 4D-CT datasets (10Hz sampling) achieved 0.12mm MAE (95%CI:0.09-0.15) for 2-sec trajectory prediction 2) GAN Respiratory Trainer with 5,000-case DIBH library generated personalized biofeedback via 4K AR interface (target: =40s breath-hold, 1.2±0.3cm displacement).
Results: -Primary Endpoints (AI-SGRT vs Conventional): - Vertical setup error: 0.07±0.06 cm vs 0.21±0.15 cm (66% reduction, p<0.001) - Rotational error: 0.53±0.41° vs 1.34±0.92° (60% reduction, p<0.001) -Secondary Endpoints:- DIBH success rate: 88% vs 62% (?+26%, p=0.01) - Treatment interruption: 3.2% vs 17.8% (OR=0.16, 95%CI:0.05-0.47) - Cardiac dose reduction: 2.3Gy mean dose (95%CI:1.8-2.7), V25%?18.6% (95%CI:15.2-22.1).
Conclusion: The AI-SGRT system achieved submillimeter motion compensation (MAE=0.12mm) and enhanced DIBH compliance by 26%, demonstrating significant potential for cardiac dose sparing. These technological breakthroughs warrant further validation through multi-institutional dose accumulation studies and SBRT applications.
Novelty Statement This first multicenter randomized controlled trial demonstrates two groundbreaking advancements in AI-enhanced radiotherapy: 1.Methodological innovation: The integration of LSTM temporal prediction (0.12mm MAE) with GAN-synthesized respiratory patterns establishes the first real-time adaptive system capable of submillimeter motion compensation in multi-site tumors. 2.Clinical impact: The 26% absolute improvement in DIBH compliance (88% vs 62%) coupled with 2.3Gy cardiac dose reduction provides Level 1 evidence for AI-driven organ sparing, surpassing existing respiratory gating technologies. -Conflict of Interest CN202422678 patent holder. No commercial funding received.