BEST IN PHYSICS (THERAPY): Intelligent In-Treatment Planning Framework for Real-Time Adaptive Radiotherapy (ART)
Abstract
Purpose
Existing adaptive radiotherapy (ART) only accounts for inter-fraction variations in anatomy. Adapted plans can become suboptimal immediately due to anatomical changes during online planning and treatment delivery, degrading treatment quality and efficacy. To address this fundamental limitation, we propose an AI-driven in-treatment adaptation framework that, for the first time, enables real-time RT adaptation.
Methods
Using prostate cancer as the testbed, in-treatment patient anatomy is continuously reconstructed from a single onboard kV image via an AI-driven real-time motion tracking module (DREME). It derives in-treatment anatomy instantly by deciphering real-time deformations from pre-treatment image. Delivered dose is accumulated in real-time via an AI-based geometry-encoding dose estimation module (GeoDose). Adaptation is automatically triggered once any large dose discrepancy is identified between delivered and planned dose. Incorporating a novel delivery-aware shuffling-convolution architecture, a real-time planning module (TransFM) is employed to adapt the plan for real-time anatomy using only the beams yet to be delivered in the ART session. To establish the proposed framework, we collected 280 patients (252 for training, 28 for validation) and tested on an independent test cohort of 20 cases using simulated motions.
Results
The end-to-end adaptation process, from image acquisition, volumetric reconstruction, dose accumulation to completion of plan update, takes ~100ms, enabling ultra-fast in-treatment adaptation to account for intra-fraction anatomical changes. Using simulated patient motion during treatment delivery with the proposed in-treatment adaptation framework, delivered dose accumulated over dynamic patient anatomy throughout the treatment is comparable to an ART plan optimized and fixed for pre-treatment patient anatomy, achieving an average γ-passing rate (3%/2mm, 10% threshold) of 99.1%±0.95% on testing cases.
Conclusion
We developed the first-of-its-kind in-treatment adaptation framework to update RT plans in real-time for instantaneous anatomy during delivery. Simulated-motion study on real patients demonstrates its potential of transforming ART to a new paradigm of dynamic in-treatment adaptation.