Towards automation of orthognathic surgery planning using deep learning for segmentation and landmark localization on cone beam computed tomography

Master thesis (2024)

Authors

B.P.P. Roumen Mechanical Engineering

Contributors

E.B. Wolvius Erasmus MC (graduation committee member)
Bart Cornelissen Erasmus MC (mentor)
Roy van den Ende Leiden University Medical Center (mentor)
Faculty
Mechanical Engineering, Mechanical Engineering
More Info
expand_more

Published Date

24-07-2024

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Mechanical Engineering

Abstract

Introduction: 3D imaging has become the standard for orthognathic surgery planning. Manual segmentation and landmark localization are the foundation for assessing deformities and planning of the corrective surgery. However, in 3D imaging, these manual procedures are complex and time consuming. Deep learning has emerged as a promising solution to automate this workflow.

Objective: To develop an automated segmentation and landmark localization workflow to reduce time and efforted required for orthognathic surgery planning using CBCT.

Methods: The dataset consisted of 57 presurgical Cone Beam CT (CBCT) scans. Manual segmentations were created for the mandible, maxilla, skin, mandibular canals (MC) left and right. Additionally, 43 landmarks (13 mandibular, 10 maxillary, 10 dental, and 10 surgical) were manually annotated. Automated landmark localization was approached as a segmentation task, with spherical segmentation of adjacent tissue with radii of 4mm, 5mm, or 6mm around landmarks. The dataset was split into 64% train, 16% validation 20% testset. Seven nnU-Net models were trained: one for segmentation and six (two models for each radius configuration) for landmark localization. Landmark positions were determined by calculating the center of mass of predicted segmentations. Performance was evaluated by comparing results to manual ground truth segmentations and landmark locations.

Results: The nnU-Net was successfully trained to identify five segmentations and 43 landmarks. The overall median [Q1 – Q3] volumetric dice coefficient (vDSC) was 0.91 [0.79 - 0.96] for the segmentation model. The overall median radial error (MRE) for the landmarking models using 4mm, 5mm and 6mm spherical segmentations (each based on two models) were 0.98 [0.58 - 1.60] mm, 1.03 [0.63 - 1.67] mm, and 1.08 [0.66 - 1.82] mm, respectively. The successful detection rate below the clinical acceptability threshold of < 2mm ranged from 78.6% to 81.6%. The inference workflow required 17.7 minutes per patient on average.

Conclusion: The nnU-Net achieved accurate craniomaxillofacial (CMF) anatomy segmentations and precise localization of landmarks, maintaining a clinical acceptability level with an error margin of less than 2mm. This fully automated workflow has the potential to enhance the efficiency of CMF planning.