Related Experiment Video
Updated: Jul 31, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
CoT-UNet++: A medical image segmentation method based on contextual transformer and dense connection
Yijun Yin1, Wenzheng Xu1, Lei Chen1
1School of Information Science and Engineering, Shandong University, Qingdao 266200, China.
Mathematical Biosciences and Engineering : MBE
|May 10, 2023
Summary
This study introduces CoT-UNet++, an improved deep learning model for segmenting individual teeth in CBCT scans. The new architecture enhances accuracy in dental diagnostics by better capturing contextual information compared to previous methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oral Health
Background:
- Accurate segmentation of individual teeth from Cone-Beam Computed Tomography (CBCT) images is crucial for diagnosing oral diseases.
- Traditional manual segmentation methods are time-consuming and labor-intensive, highlighting the need for automated solutions.
- Existing models like TransUNet, while successful, have limitations in feature fusion and context utilization.
Purpose of the Study:
- To propose a novel Context-Transformed TransUNet++ (CoT-UNet++) architecture for automated tooth segmentation in CBCT images.
- To address the limitations of TransUNet, specifically restrictive fusion and lack of adjacent key context.
- To improve the accuracy and efficiency of tooth segmentation to aid clinicians in diagnosis and treatment planning.
Main Methods:
- Developed a CoT-UNet++ architecture featuring a hybrid encoder, dense connection, and decoder.
- Employed CoTNet within the hybrid encoder to capture contextual information between adjacent keys and Transformer for global context.
- Utilized cascading upsamplers in the decoder for resolution recovery and dense concatenation for multi-scale feature fusion.
- Implemented a weighted loss function (focal, dice, cross-entropy) for pixel-level optimization.
Main Results:
- The proposed CoT-UNet++ demonstrated superior performance compared to baseline models in tooth segmentation tasks.
- The architecture effectively captured contextual information and improved the accuracy of individual tooth depiction.
- The weighted loss function contributed to reduced training error and enhanced pixel-level segmentation accuracy.
Conclusions:
- CoT-UNet++ offers a significant advancement in automated tooth segmentation from CBCT images.
- The novel architecture effectively addresses limitations of previous methods by incorporating contextual information and improved feature fusion.
- This method holds promise for enhancing the efficiency and accuracy of dental diagnostics and treatment planning.
