Abstract
In image-guided radiation therapy, extracting features from medical point cloud is the key technique for multimodality registration. This novel framework, denoted Control Point Net (CPN), provides an alternative to the common applications of manually designed keypoint descriptors for coarse point cloud registration. The CPN directly consumes a point cloud, divides it into equally spaced 3D voxels and transforms the points within each voxel into a unified feature representation through voxel feature encoding (VFE) layer. Then all volumetric representations are aggregated by Weighted Extraction Layer which selectively extracts features and synthesize into global descriptors and coordinates of control points. Utilizing global descriptors instead of local features allows the available geometrical data to be better exploited to improve the robustness and precision. Specifically, CPN unifies feature extraction and clustering into a single network, omitting time-consuming feature matching procedure. The algorithm is tested on point cloud datasets generated from CT images. Experiments and comparisons with the state-of-the-art descriptors demonstrate that CPN is highly discriminative, efficient, and robust to noise and density changes.
Highlights
1) We improve the architecture of voxel feature encoding (VFE) layer based on Voxel Net [6], which is suitable for consuming unordered point sets in 3D; 2) We propose Weighted Extraction Layer (WEL) which selectively synthesizes local features into global descriptors and coordinates of control points; 3) We show how Control Point Net (CPN) can be trained to perform 3D registration without point matching procedure; 1.2
The proposed CPN consists the following layers or functional blocks: 1) Data Grouping, in which the 3D space is subdivided into cubic voxels, points are grouped according to the voxel they reside in; 2) Random Sampling, in which fixed number inside each voxel are randomly selected; 3) Voxel Feature Encoding (VFE) Layers, in which point-wise features and locally aggregated feature are combined to learn descriptive shape; 4) Multiple Fully Connected Layers, in which each point-wise feature are aggregated within a middle layer, packed up into a fixed-length feature vector; 5) Weighted Extraction Layer, in which feature vectors in a same voxel are first normalized by softmax arithmetic along elements, synthesized to be the feature vector of this voxel
This paper presented CPN, an innovative network that extracts feature descriptors and control points for medical point cloud registration
Summary
A point cloud is a set of unorganized irregular 3D points in a unified coordinate system, capturing 3D spatial information of an object or scenery. 3D point cloud. 1) We improve the architecture of voxel feature encoding (VFE) layer based on Voxel Net [6], which is suitable for consuming unordered point sets in 3D; 2) We propose Weighted Extraction Layer (WEL) which selectively synthesizes local features into global descriptors and coordinates of control points; 3) We show how CPN can be trained to perform 3D registration without point matching procedure; 1.2. [15] represents latest progress of deep-learning-based algorithm for point cloud registration It encodes local 3D geometric structures into super-points using unsupervised auto-encoder. A matching procedure and a fine-tuning stage must be performed before transformation estimation
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
