Related Experiment Video
Updated: Jul 6, 2026
Outer-Boundary Assisted Segmentation and Quantification of Trabecular Bones by an Imagej Plugin
Published on: March 14, 2018
Automatic segmentation of the midfacial bone surface from ultrasound images using deep learning methods
1Department of Oral and Maxillofacial Surgery, Peking University School and Hospital of Stomatology, Beijing, China; National Center for Stomatology, Beijing, China; National Clinical Research Center for Oral Diseases, Beijing, China; National Engineering Research Center of Oral Biomaterials and Digital Medical Devices, Beijing, China.
Deep learning effectively segments midface bone from 2D ultrasound images. The nnU-Net algorithm achieved the highest accuracy, enabling non-invasive 3D bone modeling without radiation exposure.
Area of Science:
- Medical imaging
- Computer science
- Biomedical engineering
Background:
- Three-dimensional (3D) ultrasound technology has advanced significantly.
- Ultrasound-based 3D bone modeling, particularly for the femur, tibia, and spine, is gaining attention.
- Ultrasound offers non-invasive, radiation-free bone surface data acquisition.
Purpose of the Study:
- To develop an automatic algorithm for segmenting the midface bone surface from 2D ultrasound images.
- To utilize deep learning methods for enhancing midfacial bone segmentation accuracy.
- To compare the performance of various deep learning networks for this specific application.
Main Methods:
- Trained six deep learning networks: nnU-Net, U-Net, ConvNeXt, Mask2Former, SegFormer, and DDRNet.
- Evaluated algorithm performance using metrics like Dice coefficient (DC), intersection over union (IoU), Hausdorff distance (HD95), and average symmetric surface distance (ASSD).
- Compared automated segmentation results against ground truth data.
Main Results:
- nnU-Net demonstrated superior performance, achieving the highest Dice coefficient (89.3% ± 13.6%).
- nnU-Net also yielded the lowest average symmetric surface distance (0.11 ± 0.40 mm).
- The study confirmed nnU-Net's capability for automatic and effective midfacial bone surface segmentation.
Conclusions:
- Deep learning, specifically nnU-Net, provides an effective method for automatic midfacial bone segmentation from 2D ultrasound.
- This approach facilitates non-invasive 3D bone modeling, crucial for various medical applications.
- Further development in this area can improve diagnostic accuracy and treatment planning in craniofacial surgery.
