Academic Background and Research Motivation

In the field of dental healthcare, Cone Beam Computed Tomography (CBCT) is a widely used three-dimensional imaging technique. CBCT provides three-dimensional images of the oral cavity and is particularly effective in diagnosing odontogenic lesions. However, the segmentation of CBCT images—labeling each voxel in the image for lesions, bones, teeth, and restorative materials—is a critical yet complex task. Currently, clinical practice relies heavily on manual segmentation, which is not only time-consuming but also requires extensive expertise. To achieve automated segmentation and reduce reliance on large amounts of manually labeled data, researchers have proposed a semi-supervised learning method that incorporates oral-anatomical knowledge. This paper introduces a novel “Oral-Anatomical Knowledge-Informed Semi-Supervised Learning Model” (OAK-SSL) for 3D CBCT image segmentation and lesion detection.

Source of the Paper

This paper was co-authored by Yeonju Lee, Min Gu Kwak, Rui Qi Chen, Hao Yan, Muralidhar Mupparapu, Fleming Lure, Frank C. Setzer, and Jing Li. The authors are affiliated with institutions such as Georgia Institute of Technology, Arizona State University, University of Pennsylvania, and MS Technologies Corporation. The paper has been published in IEEE Transactions on Automation Science and Engineering and is supported by NIH Grant DE031485.

Research Process

1. Problem and Challenges

The main challenge in automating CBCT image segmentation is reducing the dependence on large amounts of manually labeled data. Existing segmentation methods typically require extensive labeled data, and manual labeling is not only time-consuming but can also lead to inconsistencies between observers. To address this issue, the researchers proposed the OAK-SSL model, which improves the segmentation accuracy of small lesions by incorporating oral-anatomical knowledge, especially when labeled data is limited.

2. Design of the OAK-SSL Model

The uniqueness of the OAK-SSL model lies in its transformation of qualitative oral-anatomical knowledge into quantitative representations and their integration into the deep learning framework. Specifically, the model includes the following three key elements: - Transformation of Knowledge into Quantitative Representation: The qualitative knowledge that “periapical lesions must be near tooth roots” is transformed into a distance map, quantifying the distance of each voxel to the nearest tooth root. - Knowledge-Informed Dual-Task Learning Architecture: The model simultaneously performs segmentation and distance prediction tasks. The distance prediction task helps the model focus on anatomically plausible lesion locations. - Knowledge-Informed Semi-Supervised Loss Function: For unlabeled images, the model leverages anatomical knowledge by combining confidence loss and stability loss to enhance the accuracy of lesion segmentation.

3. Dataset and Preprocessing

The study used a dataset of 145 3D CBCT images provided by the School of Dental Medicine at the University of Pennsylvania. The images had a resolution of 341×341×341 voxels, and all images included at least one tooth root with lesions. The samples were divided into “small lesion” and “regular lesion” groups to validate the model’s performance in segmenting small lesions. The training set consisted of 20 labeled samples and 80 unlabeled samples, while the test set included 30 labeled samples.

4. Experiments and Results

The study compared OAK-SSL with various existing methods (including supervised and semi-supervised learning methods) using metrics such as Dice scores and lesion detection accuracy. OAK-SSL performed excellently in all segmentation tasks, especially in segmenting small lesions, significantly outperforming other methods. Specifically: - Dice Score: OAK-SSL achieved a Dice score of 0.647 for lesion segmentation, significantly higher than other methods (e.g., 0.215 for supervised learning). - Lesion Detection Accuracy: OAK-SSL achieved precision and recall rates of 0.791 and 0.933, respectively, demonstrating its effectiveness in reducing false positives and false negatives.

5. Ablation Study

To verify the effectiveness of each module in OAK-SSL, the study conducted an ablation study. The results showed: - Use of Unlabeled Data: When unlabeled data was removed, the model’s Dice score for lesion segmentation dropped from 0.647 to 0.301, proving the importance of unlabeled data. - Knowledge-Informed Weights: When the knowledge-informed weights in the distance prediction task were removed, the model’s Dice score for lesion segmentation dropped from 0.647 to 0.805, indicating the critical role of these weights in enhancing model performance.

Conclusion and Value

The OAK-SSL model significantly improves the accuracy of 3D CBCT image segmentation by incorporating oral-anatomical knowledge, particularly excelling in segmenting small lesions. The model not only reduces reliance on large amounts of manually labeled data but also enhances the efficiency and reliability of the segmentation process through automation. Future research could further expand the application of the model’s anatomical knowledge, such as incorporating information on lesion shape and size and distinguishing between different types of periapical lesions to improve diagnostic and treatment planning precision.

Research Highlights

• Innovation: The first transformation of oral-anatomical knowledge into quantitative representations and their successful integration into a deep learning model.
• Practicality: Significantly reduces dependence on large amounts of manually labeled data, enhancing segmentation efficiency and accuracy.
• Clinical Value: Effectively detects and segments early-stage small lesions, providing crucial support for clinical diagnosis and treatment.

Other Valuable Information

The OAK-SSL model also utilized Gradient-weighted Class Activation Mapping (Grad-CAM) visualization technology to demonstrate the model’s decision-making process in lesion segmentation. The results showed that OAK-SSL can more accurately focus on actual lesion areas, avoiding the false-positive issues commonly seen in other methods.