Surface-Guided 3D Liver Tumor Tracking for Respiratory Motion Management Using Deep Learning
Abstract
Purpose
Respiratory motion management remains a challenge for radiotherapy in abdominal regions. This study aims to develop and validate a robust framework that correlates external body surface motion with internal liver tumor movement, enabling dose-free, real-time 3D liver tumor tracking based solely on surface.
Methods
A dataset of 157 patients with liver tumors undergoing 4D MRI was collected. Tumor contours were transferred from paired planning CTs, and body surface contours were extracted from 4D MRIs (eight phases). A 3D U-Net–based model was developed to predict the real-time 3D tumor mask at each respiratory phase using synchronized surface combined with anatomical priors from a reference end-exhalation phase and surface deformation features. A patient-adaptive input-attention mechanism was introduced to learn spatially varying surface importances. Surface significance was investigated through both physical correlation analysis and learned attention maps. An uncertainty-aware extension incorporating aleatoric uncertainty estimation was developed to provide case-wise confidence scores.
Results
The model accurately tracked liver tumor motion (tumor centroid error 1.89 ± 1.40 mm, dice score 0.81 ± 0.16). Augmented inputs—including diaphragm masks, tumor-highlighted surfaces, and surface deformable registration–derived features—improved tracking performance. Tumors with larger respiratory motion amplitudes exhibited increased localization errors. Surface analysis identified the lower chest, abdominal, lateral body, and regions proximal to the tumor as the most informative for motion prediction, with patient-specific spatial patterns captured by the learned attention maps. The model also produced aleatoric uncertainty estimates, with higher variance correlating with reduced prediction accuracy (p=7.46×10-6).
Conclusion
This study demonstrates the feasibility of surface-guided internal tumor localization for liver radiotherapy, offering a promising pathway toward real-time, radiation-free motion management. The developed patient-adaptive attention mechanisms allow the model to focus on clinically informative surface regions. The integration of uncertainty estimation further provides a quantitative measure of model reliability, supporting safer and more informed clinical deployment.