Efficient and Robust CBCT Segmentation of Oral and Maxillofacial Structures

Fan Xiao, Xinrui Huang, Anqi Gao, Dongming He, Xiaofan Zhang, Xudong Wang

Keywords: CBCT image, Oral and maxillofacial structures segmentation, Interactive segmentation

Abstract: In dental practice, accurate segmentation of oral and maxillofacial structures from cone-beam computed tomography (CBCT) images is essential for diagnostic and treatment planning purposes. However, manual segmentation is time-consuming and labor-intensive. Although numerous deep learning-based methods have been proposed to automate this process, most rely on a single model architecture, which struggles to handle the complex and diverse nature of oral anatomical structures. To address this limitation, we propose a hybrid framework integrating nnUNet and VISTA models for automated and interactive segmentation of oral and maxillofacial structures. Our approach employs a class-wise ensemble strategy to improve inference efficiency and accuracy, and incorporates post-processing techniques such as threshold-based small object removal and disconnected region filtering to enhance robustness. The proposed method achieved third place in Task 1 and second place in Task 2 of the ToothFairy3 Challenge. Code and model weights are available at \url{https://github.com/ff741333/toothfairy3_blcakmyth}.

Changes Summary:

For Reviewer 4Wio,
p.1: We have now explicitly added the GitHub repository URL in the text for better visibility, in addition to the embedded link in the PDF.
p.2: The phrase “Typically, the oral and maxillofacial structures…” has been revised to “The oral and maxillofacial structures…” as suggested.
p.3: The redundant phrase “For the interactive segmentation task,” at the beginning of the sentence has been removed.
Sec. 2 Method: We have expanded the figure 2 of how VISTA without prompt and VISTA with prompt are integrated during inference.
Augmentation hyperparameters have been added in Section 2.1.
Sec. 2.4 Postprocessing: We have provided a more detailed explanation of the postprocessing steps, including the hyperparameters.
p.4: The term “Poly scheduler” now includes a reference. And for specific methods, please refer to https://github.com/ff741333/toothfairy3_blcakmyth/blob/main/task1/train/nnUNet/nnunetv2/training/lr_scheduler/polylr.py
p.5: The typo “VISIT model” has been corrected to “VISTA model”.
p.6: “dice score” has been changed to “Dice score” throughout the text.
p.6: Remove the "both"

For Reviewer f6eo,
Citation of Challenge Dataset: Add citation at Sec. 2.1
GitHub Links: We have now explicitly added the GitHub repository URL in the text for better visibility, in addition to the embedded link in the PDF.
Citation Spacing: We have performed a thorough proofreading pass throughout the manuscript.
Updated Challenge Ranking: The section describing our algorithm's placement has been revised to reflect the official final ranking from the ODIN challenge.
Running Title: We have provided a concise running title to replace the suppressed header. The new running title is: "Efficient and Robust CBCT Segmentation".
Consistent Spelling of VISTA: The typo “VISIT model” has been corrected to “VISTA model”.

Main Tex File: main.tex

Confirm Latex Only: true

Code Url: https://github.com/ff741333/toothfairy3_blcakmyth

Authors Changed: false

Submission Number: 16