Intra-fractional Voxel-wise Anatomical Motion Tracking Guided by Multimodal Respiratory Surrogates in Radiotherapy: Framework Development and Multi-Center Validation.

Abstract

PurposeIntra-fractional respiratory motion management encounters trade-offs between target coverage, organ at risk (OAR) sparing, and treatment efficacy. Real-time tracking is promising but hindered by invasiveness risks, lack of global deformations, or inadaptability to irregular respiration. Currently respiratory modeling-based adaptive tracking may offer ideal solutions, but is limited by discrete clinical priors, highly-sparse observations, and complex motion variability. This study advances the constrained modeling toward a synergistic knowledge- and information-augmentation-driven paradigm (KIADP), thereby bridging sparse real-time observations with voxel-wise anatomical deformations.Methods and materialsFour multi-center cohorts were enrolled comprising 35 cases (10 prospective and 25 retrospective), including 7 patients undergoing re-irradiation. KIADP was instantiated as MorphTracking, a patient-specific respiratory modeling framework that synergistically leveraged prior clinical materials and external-and-internal surrogates-optical surface images (OSIs) and single-view X-ray projections-to estimate 3D deformation vector fields (DVFs). MorphTracking incorporated a comprehensive data augmentation pipeline to enhance phase diversity with irregular variability. To enforce physical-consistency and anatomical-plausibility, we performed optimization spanning deformation, image, and surrogate domains. MorphTracking was compared with three state-of-the-art (SOTA) methods and further validated through full-cycle evaluations, ablation studies, and simulated 4D treatment scenarios.ResultsAcross thoracic (n=24) and abdominal (n=11) cohorts, MorphTracking significantly outperformed SOTA methods in CT reconstruction and tracking of gross tumor volume (GTV) and OARs. It exhibited high consistency with ground-truths spanning full-cycle phases, operating at an average inference time of 15.622±0.001 ms/frame. Comprehensive ablation studies substantiated the individual and synergistic contributions of data augmentation, multimodal configuration, and cross-domain optimization. In simulated treatment scenarios, compared to the standard internal GTV (IGTV) approach, MorphTracking-driven 4D delivery effectively maintained GTV's dosimetric indices while enhancing OAR (spinal cord and lungs) sparing.ConclusionsMorphTracking provides an innovative solution for intra-fractional respiratory modeling, potentially enhancing voxel-wise geometric confidence for motion-aware highly-conformal dose delivery with real-time, robust, and adaptive anatomical tracking.