Semi-supervised CBCT–IOS Registration Using PointNetLK

AJO BABU GEORGE, GADHA LEKSHMI P, SADHVIK BATHINI, Govind A

Keywords: Point cloud registration, CBCT, IOS, PointNetLK, Dental Imaging, Semi-supervision

TL;DR: Accurate alignment of CBCT-IOS multimodal data using semi-supervised learning for unified PointNetLK architecture.

Abstract: Accurate alignment of intraoral scans (IOS) with cone-beam computed tomography (CBCT) is essential for integrated dental diagnostics and surgical planning. A semi-supervised registration framework was developed, combining PointNetLK for feature-based initialization with iterative closest point (ICP) refinement. Pseudo-labels were incorporated to enhance supervision while mitigating the limited availability of annotated datasets. Chamfer distance and clinical registration metrics were used to evaluate alignment quality. Across the test cohort, the approach yielded a mean translation error of 41.67 mm and a mean rotation error of 33.96°, highlighting the challenge of partial-arch fusion. Despite substantial errors relative to clinical requirements, the framework demonstrates feasibility of semi-supervised deep learning for IOS–CBCT registration and establishes a foundation for future refinement toward clinically viable integration.

Changes Summary: We addressed reviewer concerns by clarifying pseudo-label selection, analyzing clinical error gaps, and discussing the limitations of the fixed HU threshold. Figures and citation formatting were improved. Three relevant references on semi-supervised segmentation and dataset limitations were added. The revised version enhances transparency, robustness, and alignment with clinical deployment goals. We have also included the necessary citations as recommended by the organizers.

Main Tex File: paper.tex

Confirm Latex Only: true

Code Url: https://github.com/Ajogeorge29/STS_MCCAI_TASK02

Dataset Url: https://www.codabench.org/competitions/6470/#/pages-tab

Authors Changed: false

Submission Number: 24