Geometry-Aware Semi-Supervised 4-D Point Cloud Reconstruction for Dynamic MRI/CT Imaging

Dynamic medical imaging has deep impact in modern healthcare by enabling visualization of anatomical structures. Four-dimensional magnetic resonance imaging (4D-MRI) and four-dimensional computed tomography (4D-CT) are primarily used in medical applications such as lung tumor tracking, cardiac motion analysis, and planning of adaptive radiotherapy. However, conventional voxel-based image reconstruction methods experience high computational complexity, motion artifacts, and limited capability to preserve temporal consistency during dynamic reconstruction. To address these limitations, this study proposes a semi-supervised 4-D point cloud reconstruction framework for dynamic MRI/CT imaging. The proposed method transforms 3-D image sequences into spatio-temporal point cloud representations. 4-D images efficiently capture anatomical geometry and temporal motion. A semi-supervised learning strategy is used to utilize both labeled and unlabeled temporal phases, and temporal consistency constraints are introduced to maintain stable anatomical representations across respiratory cycles. The framework is evaluated using the TCIA 4D Lung dataset. Comparative analysis demonstrates improvements in reconstruction quality, geometric fidelity, and temporal consistency compared with conventional voxel-based approaches. The proposed framework provides a geometry-aware and computationally efficient alternative for dynamic medical image reconstruction and offers significant potential for future motion-aware diagnostic and therapeutic applications.