In image-guided radiation therapy, extracting features from medical point\ncloud is the key technique for multimodality registration. This novel framework,\ndenoted Control Point Net (CPN), provides an alternative to the common\napplications of manually designed key-point descriptors for coarse point\ncloud registration. The CPN directly consumes a point cloud, divides it into\nequally spaced 3D voxels and transforms the points within each voxel into a\nunified feature representation through voxel feature encoding (VFE) layer.\nThen all volumetric representations are aggregated by Weighted Extraction\nLayer which selectively extracts features and synthesize into global descriptors\nand coordinates of control points. Utilizing global descriptors instead of local\nfeatures allows the available geometrical data to be better exploited to improve\nthe robustness and precision. Specifically, CPN unifies feature extraction\nand clustering into a single network, omitting time-consuming feature\nmatching procedure. The algorithm is tested on point cloud datasets generated\nfrom CT images. Experiments and comparisons with the state-of-the-art\ndescriptors demonstrate that CPN is highly discriminative, efficient, and robust\nto noise and density changes.
Loading....