Semi-Supervised 4-D Point Cloud Reconstruction for Motion-Consistent Dynamic MRI/CT Imaging

Dynamic medical imaging techniques, such as four-dimensional magnetic resonance imaging (4D-MRI) and four-dimensional computed tomography (4D-CT), are used in applications such as lung tumor tracking, cardiac motion analysis and vascular flow modeling and allow physicians to visualize the motion of organs over time. However, the conventional voxel-based reconstruction techniques are computationally intensive and they fail to provide the accurate description of temporal correspondence and geometric features in dynamic medical images.

This research introduces a semiautomatic 4-D point cloud reconstruction system of dynamic MRI/CT images. The proposed solution incorporates semi-supervised deep learning to utilize the identified and the labels of temporal frames as well as transforms volumetric medical images into the form of spatio-temporal point clouds. Temporal consistency requirements are used to maintain anatomical stability during respiratory stages. Experiments on the TCIA 4D Lung dataset reveal that when compared to voxel-based reconstruction methods, the reconstruction accuracy, geometric fidelity and temporal consistency are improved. It has been found that efficient geometry-based medical image reconstruction framework is provided by point-cloud-based representations.