An Artificial Intelligence Hypothetical Approach for Masseter Muscle Segmentation on Ultrasonography in Patients With Bruxism

Abstract:

Aim

The study aims to assess the success of an AI system based on the deep convolutional neural network (D-CNN) for segmenting masseter muscles on ultrasonography (USG) images.

Materials and Methods

• Study Design: Retrospective study using 195 anonymized USG images from the radiology archive of the Faculty of Dentistry, Ankara University.
• Technology: Utilized U-net, Pyramid Scene Parsing Network (PSPNet), and Fuzzy Petri Net (FPN) architectures for deep learning.
• Comparison: Manual segmentation and measurements were statistically compared with AI results.
• Statistical Analysis: Accuracy, ROC AUC, and PRC AUC were calculated and compared between human observers and the AI model using the Mann–Whitney U test.

Results

• AI models (FPN, PSPNet, U-net) demonstrated high accuracy in detecting and segmenting masseter muscles.
• The D-CNN measurements were consistent with manual measurements, showing no significant differences.

Conclusion: The AI system is promising for automatic masseter muscle segmentation and thickness measurement on USG images, potentially aiding professionals in diagnosis and saving time.

Introduction

• Bruxism, affecting 8% to 31% of adults, leads to masseter muscle issues such as inflammation and hypertrophy.
• Ultrasonography (USG) is a key diagnostic tool for muscle assessment, but variability in results due to operator experience is a challenge.
• AI and deep learning offer potential improvements in accuracy and efficiency for medical imaging.

Materials and Methods:

Data Collection

• USG images of 24 patients diagnosed with bruxism were collected.
• A total of 195 images were used, divided into training (157), validation (18), and test (20) groups.

Image Annotation

• Manual annotations by an experienced radiologist were used as ground truth.
• AI architectures employed included U-net, PSPNet, and FPN.

Model Training and Evaluation

• AI models were trained to detect and segment masseter muscles.
• Performance metrics such as accuracy, sensitivity, specificity, precision, F1-score, ROC AUC, and PRC AUC were calculated.

Results

• The AI models, particularly FPN and U-net, showed high accuracy (0.985 and 0.969 respectively).
• No significant difference was found between AI and manual measurements (P > .05).

Discussion

• AI systems show potential in improving diagnostic accuracy and efficiency for masseter muscle assessment in bruxism patients.
• The findings align with previous research on AI's effectiveness in medical imaging.

Conclusion

The AI-based approach for USG image analysis is effective for automatic masseter muscle segmentation, aiding in diagnosis and potentially enhancing clinical workflows.