Bayesian model-based liver respiration motion prediction and evaluation using single-cycle and double-cycle 4D CT images

Abstract

To reduce the effects of respiratory movements in abdominal organs has been a complex issue in radiation therapy. The study aimed to introduce the use of machine learning methods to construct patient-specific respiratory motion models. A Bayesian-based PCA statistical model was proposed and 6 patients were used as experimental data. The correct rate of PCA statistical model estimation follows a probability distribution with respect to model parameters. Combined with Bayesian probabilistic reasoning, the preoperative statistical model is used to estimate the prior probability, and the likelihood ratio is constructed according to the similarity between intraoperative ventral surface and preoperative CT surface. Therefore, the posterior probability of the current internal respiratory motion vector field can be obtained. By maximizing the posterior probability, the optimal PCA statistical model parameters can be obtained, and then the internal respiratory motion estimation with maximum posterior probability can be obtained. To validate the motion estimation accuracy of the respiratory motion model, we used abdominal 4D CT images of 6 cases for construction and testing. For each set of abdominal 4D CT images, the abdominal respiratory motion vector field (DVF) was calculated after determining the reference phase, and the abdominal CT surface was extracted. In this paper, when using single-cycle CT data, for a statistical motion model with Bayesian inference, the average error of motion estimation is 0.57±0.06 mm. When using experimental two-cycle CT data, the average error of motion estimation is 1.52±0.41 mm. Preliminary experimental results show that the model obtained similar motion estimation errors comparable with state-of-the-art.