Self-Configuring 3D Segmentation of Pediatric Dentition

Enzo Tulissi, Alban Gaydamour, Juan C. Prieto, Claudia Mattos, Sara Tinawi, Aron Aliaga Del Castillo, Dylan J. Keener, Marilia Yatabe, Eduardo Caleme, Brent Larson, Antonio C. de Oliveira Ruella, Jonas Bianchi, Heesoo Oh, Marcela Lima Gurgel, Erika Benavides, Fabiana Soki, Gustavo A. Ruíz-Mora, Yalil A. Rodríguez-Cárdenas, Bruno M. R. Braga, Ana B. Teodoro et al. (4 additional authors not shown)

Keywords: 3D segmentation, pediatric dentistry, Cone-beam CT (CBCT), deep learning, nnU-Net, dental morphometry

TL;DR: This paper presents a self-configuring deep learning approach with nnU-Net v2 for accurate 3D segmentation and universal labeling of primary and permanent teeth in pediatric CBCT scans, improving dental morphometry and clinical applications.

Abstract:

Robust and anatomically accurate 3D segmentation of both primary and permanent teeth in cone-beam computed tomography (CBCT) is a foundational task in pediatric
and orthodontic diagnosis. Precise segmentation not only facilitates virtual modeling and morphometric analysis but also supports shape-aware clinical decision-making for mixed dentition.
We present a fully automatic deep learning approach based on the self-configuring nnU-Net v2
framework, optimized for geometric modeling and shape-based interpretation of dental structures. Our method leverages the volumetric architecture of nnU-Net to learn and delineate
fine-scale dental geometries without the need for hand-crafted tuning. Evaluated on a large
dataset of pediatric CBCTs, with 369 scans for training, 93 for validation, and 55 for testing,
the model achieves high accuracy across all teeth, with mean Dice score 0.87. We detail our
preprocessing pipeline, volumetric architecture, data augmentation, and postprocessing steps designed to preserve anatomical fidelity. Results demonstrate that our framework
generates consistent, clinically actionable 3D segmentations, enabling advanced shape analy-
sis and digital treatment planning in pediatric dentistry. This work contributes to the field of
geometric learning in medical imaging by extending state-of-the-art volumetric segmentation
to mixed dentition CBCT data, and supports the integration of AI into clinical
workflows.

Changes Summary: I’ve included additional figures and refined their descriptions to address the reviewers’ feedback.

Main Tex File: main.tex

Confirm Latex Only: true

Code Url: https://github.com/ashmoy/SlicerAutomatedDentalTools/tree/main/BATCHDENTALSEG

Authors Changed: true

Submission Number: 3