Author: Yunxiang Li, Hua-Chieh Shao, Chenyang Shen, Jing Wang, Jiacheng Xie, Shunyu Yan, You Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, The University of Texas at Dallas 🌍
Purpose: Accurate liver deformable motion tracking is essential in image-guided radiotherapy to enable precise tumor targeting during treatment. We developed a conditional point cloud diffusion model for accurate deformable liver motion estimation and tumor localization, based on an arbitrarily-angled X-ray projection.
Methods: PCD-Liver estimates deformable liver motion by solving deformable vector fields (DVFs) of a prior liver surface point cloud, based on a single X-ray image. It is a patient-specific model of two main components: a rigid alignment model and a conditional point cloud diffusion model. The rigid alignment model estimates the liver’s overall shifts, and the point cloud diffusion model further corrects for the liver surface’s local deformation. Conditioned on the motion-encoded features extracted from a single X-ray projection by a geometry-informed feature pooling layer, the diffusion model iteratively solves detailed liver surface DVFs in an angle-agnostic fashion. The liver surface motion solved by PCD-Liver is subsequently fed as the boundary condition into a UNet-based biomechanical model to infer the liver’s internal motion to localize liver tumors.
A dataset of 10 liver cancer patients was used for evaluation. Data augmentations were implemented with a principal component analysis-based motion model to simulate realistic, respiration-induced liver motion, yielding 3072/1536/135 volumes for training/validation/testing. We used the root-mean-square-error (RMSE) and 95-percentile Hausdorff distance (HD95) metrics to examine the liver point cloud motion estimation accuracy, and the center-of-mass-error (COME) to quantify the liver tumor localization error.
Results: PCD-Liver outperformed a state-of-the-art, graph neural network-based X360 model. The mean(±s.d.) RMSE, HD95, and COME of the prior liver or tumor were 8.86mm(±1.51mm), 10.88mm(±2.56mm), and 9.41mm(±3.08mm), respectively. After PCD-Liver’s motion estimation, the corresponding values were 3.59mm(±0.28mm), 4.29mm(±0.62mm), and 3.45mm(±0.96mm). In comparison, the values were 4.63mm(±0.29mm), 5.49mm(±0.47mm), and 3.98mm(±0.94mm) for X360.
Conclusion: PCD-Liver estimates liver motion accurately from a single, arbitrarily-angled X-ray projection, allowing real-time precise tumor localization.