nnUNet for Semi-supervised Tooth and Pulp Root Canal Segmentation in CBCT

AJO BABU GEORGE, SADHVIK BATHINI

Keywords: Teeth Segmentation · Semi-supervised learning · nnUNet · CBCT · Pulp · Root canals

TL;DR: A semi-supervised nnU-Net model that leverages a small labeled and a large unlabeled datasets and accurately segments teeth and pulp root canals in 3D CBCT scans.

Abstract: A solution for the Semi-supervised Teeth Segmentation and Registration (STSR) 2025 Challenge, which focused on the precise segmentation of teeth and pulp root canals in 3D Cone Beam Computed Tomography (CBCT) scans is presented in this paper. Accurate segmentation of the pulp root canal is crucial for clinical visualization and treatment planning, but manual annotation is extremely labor-intensive. The presented approach uses a semi-supervised framework powered by nnU-Net, leveraging a small labeled dataset of 30 scans alongside a much larger unlabeled dataset of 300 scans. To effectively utilize the unlabeled data, pseudo-labeling was employed to generate annotations, and the model was subsequently trained. The results demonstrate strong segmentation performance for both tooth and pulp structures, achieving a Dice score of 0.8088 and an mIoU of 0.9638. These metrics highlight the model's ability to accurately identify and delineate the target structures.

Changes Summary: We have addressed all reviewer comments and updated the manuscript accordingly. Metric reporting inconsistencies were resolved by linking Tables 4 and 5 to anatomical labels and evaluation setup. Data augmentation strategies—including flipping, rotation, and noise injection—are now detailed in the training section. The pseudo-labeling procedure was clarified with a threshold-based high-confidence selection strategy. A new “Limitations and Future Work” section discusses clinical deployment gaps, including missed apical segmentation and plans for active learning. Formatting, figures, and terminology were polished throughout. Finally, all three mandatory STSR challenge citations have been properly included.

Main Tex File: paper.tex

Confirm Latex Only: true

Code Url: https://github.com/Ajogeorge29/STS_MCCAI_TASK_01

Dataset Url: https://www.codabench.org/competitions/6468/

Authors Changed: false

Submission Number: 23