The patient-specific anatomy and kinematics are important for preoperative planning, rehabilitation, biomechanical simulation, and so on. In this presentation, I discuss followings: 1) Recovery of three-dimensional rib motion from dynamic chest radiograph (x-ray video). 2) Automatic segmentation of hip and thigh muscles from CT and MR Images using convolutional neural networks (CNN) and image synthesis technique.
Recovery of three-dimensional rib motion: Previous analysis of rib-cage mo- tion using x-ray video has been shown to be effective for evaluating respiratory function, but it has been limited to 2D. I focus on a method recovering 3D rib motion based on 2D-3D registration of x-ray video and single-time-phase CT. I introduce the following two components into the conventional intensity-based 2D–3D registration pipeline: i) a rib-motion model based on a uniaxial joint to constrain the search space and ii) local contrast normalization as a pre-process of x-ray video to improve the cost function of the optimization parameters, which is often called the landscape.
Automatic segmentation of hip and thigh muscles from CT and MR: Previous automated segmentation from CT, based on a hierarchical multi-atlas method, has been shown to be high accuracy, but it has been limited to low computational efficiency. I focus on segmentation from CT using CNN, which yielded significant improvement of accuracy and computational time. Then, I focus on MR-to-CT synthesis to realize modality independent segmentation. I extend the CycleGAN approach by adding the gradient consistency loss to improve the accuracy at the boundaries. Finally, I introduce an active-learning method using Bayesian approach to effectively increase training dataset size.