Real-Time Deformable Liver Motion Tracking By a Single Arbitrarily-Angled X-Ray Projection Via a Latent Diffusion Model (Latent-Liver)
Abstract
Purpose
Real-time liver motion tracking is essential in image-guided radiotherapy to enable precise tumor targeting. We developed a conditional latent point cloud diffusion model (Latent-Liver) for real-time deformable liver motion tracking and tumor localization using an arbitrarily-angled X-ray projection.
Methods
From a single X-ray projection, Latent-Liver estimates 3D liver surface motion by predicting the deformation-vector-field (DVF) that warps a prior liver surface point cloud to the target surface. This patient-specific model runs sequentially: (1) conditioned on feature vectors extracted from an arbitrarily-angled X-ray projection via a geometry-informed feature pooling layer, a latent diffusion model iteratively denoises a noise token to generate a latent DVF code, (2) a latent compressor/decoder converts this latent code into the full DVF, and (3) a rigid alignment model estimates the liver’s global rigid motion, which is then composed with the DVF to produce the final liver surface motion field. The resulting surface motion is then applied as a boundary condition to a U-Net-based biomechanical model to infer internal deformation for liver tumor localization. We evaluated Latent-Liver on 7 liver cancer patients. A principal component analysis(PCA)-based respiratory motion model was used for augmentation, generating 3072/1536/135 volumes for training/validation/testing per patient. Liver point cloud motion accuracy was assessed using root-mean-square-error(RMSE) and 95-percentile Hausdorff distance(HD95), and tumor localization error was quantified by center-of-mass-error(COME).
Results
Compared with a state-of-the-art, point cloud diffusion-based model (PCD-Liver), Latent-Liver achieves 200x faster (450ms vs 90s) inference while retaining motion estimation accuracy. The mean(±s.d.) RMSE, HD95, and COME of the prior liver or tumor (before motion estimation) were 9.07mm(±3.66mm), 11.20mm(±4.83mm), and 9.90mm(±5.39mm), respectively. After Latent-Liver’s motion estimation, the corresponding values were 4.13mm(±1.99mm), 4.56mm(±2.05mm), and 3.96mm(±2.42mm). In comparison, the values were 3.60mm(±1.92mm), 4.07mm(±1.81mm), and 3.70mm(±2.32mm) for PCD-Liver.
Conclusion
Latent-Liver estimates liver motion accurately and efficiently from an arbitrarily-angled X-ray projection, allowing real-time precise tumor localization.