Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scaling nnU-Net for CBCT Segmentation

Fabian Isensee, Yannick Kirchhoff, Lars Kraemer, Maximilian Rokuss, Constantin Ulrich, Klaus H. Maier-Hein

TL;DR

This work addresses multi-structure segmentation in dental CBCT via a tailored nnU-Net ResEnc L configuration. The authors increase patch size, deepen the network, adjust augmentation (notably disabling left-right mirroring), extend training, and apply per-class postprocessing cutoffs, achieving a test Dice of $0.9253$ and HD95 of $18.472$, with a mean rank of $4.6$ and first place in ToothFairy2. They validate gains through five-fold cross-validation and qualitative analyses, and provide open-source code for reproducibility. The approach demonstrates robust, high-precision CBCT segmentation relevant for dental diagnostics and planning, with practical implications for automated clinical workflows.

Abstract

This paper presents our approach to scaling the nnU-Net framework for multi-structure segmentation on Cone Beam Computed Tomography (CBCT) images, specifically in the scope of the ToothFairy2 Challenge. We leveraged the nnU-Net ResEnc L model, introducing key modifications to patch size, network topology, and data augmentation strategies to address the unique challenges of dental CBCT imaging. Our method achieved a mean Dice coefficient of 0.9253 and HD95 of 18.472 on the test set, securing a mean rank of 4.6 and with it the first place in the ToothFairy2 challenge. The source code is publicly available, encouraging further research and development in the field.

Scaling nnU-Net for CBCT Segmentation

TL;DR

This work addresses multi-structure segmentation in dental CBCT via a tailored nnU-Net ResEnc L configuration. The authors increase patch size, deepen the network, adjust augmentation (notably disabling left-right mirroring), extend training, and apply per-class postprocessing cutoffs, achieving a test Dice of and HD95 of , with a mean rank of and first place in ToothFairy2. They validate gains through five-fold cross-validation and qualitative analyses, and provide open-source code for reproducibility. The approach demonstrates robust, high-precision CBCT segmentation relevant for dental diagnostics and planning, with practical implications for automated clinical workflows.

Abstract

This paper presents our approach to scaling the nnU-Net framework for multi-structure segmentation on Cone Beam Computed Tomography (CBCT) images, specifically in the scope of the ToothFairy2 Challenge. We leveraged the nnU-Net ResEnc L model, introducing key modifications to patch size, network topology, and data augmentation strategies to address the unique challenges of dental CBCT imaging. Our method achieved a mean Dice coefficient of 0.9253 and HD95 of 18.472 on the test set, securing a mean rank of 4.6 and with it the first place in the ToothFairy2 challenge. The source code is publicly available, encouraging further research and development in the field.

Paper Structure

This paper contains 12 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Baseline configuration
  4. Our improvements
  5. Things we tried that didn't work
  6. Results
  7. Quantitative Results (Training Set)
  8. Qualitative Results (Training Set)
  9. Test Set Submission
  10. Quantitative Results (Test Set)
  11. Conclusion
  12. Acknowledgments.

Figures (2)

  • Figure 1: Qualitative Results of our method (single model), predictions are obtained from the validation sets of each fold in our 5-fold cross-validation. Ground Truth is shown on the left with both, a 3D rendering and a representative 2D slice. Correspondingly, the predictions are shown on the right. Most cases are predicted accurately and all teeth are correctly classified (top). Small errors include misclassification of teeth or inconsistencies between crowns and teeth (middle). Severe artifacts can cause poor performance (bottom).
  • Figure 2: Illustration of our method's robust performance on challenging cases, including artifacts, noise, and anatomical outliers. Again the predictions are obtained from the validation sets of our 5-fold cross-validation.