Real-time Image-guided Radiotherapy via Joint Target Localization and Volumetric Reconstruction

Radiation therapy is a critical treatment modality for malignant tumors, aiming to deliver a precise, spatially conformal radiation dose to the tumor while sparing the organs-at-risk (OARs). However, due to respiratory, gastrointestinal and urinary activity, internal patient anatomy undergoes motion which can compromise the accuracy of radiotherapy. To address this challenge, we propose a patient-specific deep learning (DL) framework capable of localizing tumors and OARs on X-ray digital radiographs (DRs) and simultaneously reconstructing volumes from single-view DRs in real-time. This framework first constructs a motion model based on the patient’s 4DCT image sequence, incorporating both global deformation and inter-organ interactions to accurately capture respiratory-induced anatomical variations. Next, LocatorNet, a rapid X-ray localization network, is pre-trained to efficiently determine spatial positioning. Guided by LocatorNet, CrossReconFormer is introduced, which integrates a high-performance 2D/3D cross-modal fusion module. This model effectively reconstructs high-fidelity CT images from single-view X-ray projections, preserving structural details and anatomical consistency. During treatment, real-time X-ray DRs are fed into the model, which yields both target localization and reconstructed CT images. Rigorous experiments conducted on patient pancreas and lung data achieve high accuracy and promising results. Clinically relevant dose calculation based on the reconstructed images indicates that our method holds significant clinical potential for accurate dose delivery, advancing the integration of DL in radiotherapy for enhanced patient care and treatment outcomes.