Related Experiment Video
Updated: Jun 2, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Deep learning for automatic segmentation of the inferior alveolar nerve using a hybrid CNN-transformer framework
Ho-Kyung Lim1, Seok-Ki Jung2, Yongwon Cho3
1Department of Oral and Maxillofacial Surgery, Korea University Guro Hospital, Seoul, 08308, Korea.
Scientific Reports
|May 31, 2026
Summary
This study introduces an improved AI framework for automatically identifying the inferior alveolar nerve (IAN) in dental scans. The new method enhances segmentation accuracy and boundary continuity, crucial for preventing nerve injury during oral surgeries.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oral and Maxillofacial Surgery
Background:
- Accurate identification of the inferior alveolar nerve (IAN) is critical for preventing nerve damage in dental and maxillofacial procedures.
- Manual annotation of the IAN in cone-beam computed tomography (CBCT) is difficult due to image noise, anatomical variations, and complex nerve pathways.
- Existing automated methods require improvement in robustness and anatomical continuity for thin, tubular structures like the IAN.
Purpose of the Study:
- To develop and evaluate an improved automatic segmentation framework for IAN identification in CBCT images.
- To enhance the robustness and anatomical continuity of IAN segmentation using a hybrid CNN-attention architecture.
Main Methods:
- Proposed a novel hybrid CNN-attention architecture built upon the nnU-Net framework for IAN segmentation.
- Incorporated stage-restricted Permuted Adaptive Instance Normalization (Permuted AdaIN) in the encoder for appearance invariance.
- Utilized a decoder-stage contextual refinement module to improve segmentation consistency of the thin tubular nerve structure.
Main Results:
- The proposed method achieved Dice similarity coefficients of 0.63 ± 0.17 (internal) and 0.62 ± 0.12 (external), showing modest improvement over nnU-Net.
- Lower boundary errors were observed: HD95 values of 2.96 ± 1.27 mm (internal) and 3.72 ± 7.63 mm (external).
- Qualitative analysis indicated improved boundary alignment and continuity of the segmented IAN trajectory.
Conclusions:
- The proposed hybrid CNN-attention framework demonstrates potential for improving automated IAN segmentation consistency and boundary agreement in CBCT images.
- The method shows promise in addressing challenges related to image noise and anatomical variability in IAN identification.
- Further validation is needed to confirm the clinical significance of the observed improvements over existing frameworks like nnU-Net.