A Deep Learning Framework for Markerless Real-Time Image-Guided Liver Tumor Radiotherapy on Conventional Linear Accelerators
Abstract
Purpose
During liver radiotherapy, respiratory-induced tumor motion can result in target underdose and increased irradiation of surrounding healthy tissue. Existing real-time image-guided approaches often require specialized equipment or invasive marker implantation due to poor soft-tissue contrast in X-ray imaging. This study proposes a markerless deep learning framework that predicts tumor position from kilovoltage (kV) X-ray images on conventional linear accelerators. A multi-institutional, multi-platform dataset is used to demonstrate the feasibility of the system.
Methods
A patient-specific conditional Generative Adversarial Network (cGAN) was trained using digitally reconstructed radiographs generated from augmented planning CT datasets. Testing was performed on kV X-ray images acquired from pre-treatment CBCT projections, which resemble intra-treatment fluoroscopic images. Projections from 97 fractions across 24 liver cancer patients, including breath-hold and free-breathing modalities, were included from the prospective multi-institutional LARK trial (NCT02984566). Ground-truth tumor positions were inferred from fiducial markers, which were masked in training and testing datasets to emulate a markerless workflow. This study hypothesized that the framework would meet clinically relevant localization errors within 5 mm of ground-truth for 95% of observations per patient. System latency was assessed against AAPM Task Group 264 (TG-264) real-time image-guided radiotherapy limit (< 500 ms).
Results
For 21 of 24 patients, 95% of AP/LR localization errors were within 5 mm, while 95% of SI errors were within 5 mm for 17 of 24 patients. End-to-end system latency was 189 ± 29 ms, meeting TG-264 limits. Across patients, the mean localization errors were -0.2 ± 0.7 mm in the AP/LR direction and -1.0 ± 1.8 mm in the SI direction.
Conclusion
Our cGAN achieved clinically acceptable localization for most patients with latency compatible with TG-264 limits. These findings demonstrate that standard kV X-ray images can be leveraged for markerless liver tumor localization, supporting future non-invasive, real-time adaptive radiotherapy on conventional linear accelerators.