Fully automated condyle segmentation using 3D convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oral and Maxillofacial Surgery

Background:

• Accurate segmentation of the mandibular condyle is crucial for diagnosing and treating temporomandibular joint disorders.
• Manual segmentation is time-consuming and prone to inter-observer variability.
• Deep learning, specifically 3D U-Net, offers a promising approach for automated medical image segmentation.

Purpose of the Study:

• To develop and evaluate an automated segmentation algorithm for mandibular condyles using 3D U-Net.
• To determine the optimal dataset size for achieving clinically acceptable segmentation accuracy through a stress test.
• To compare the performance of a basic 3D U-Net with a cascaded 3D U-Net architecture.

Main Methods:

• Acquired 234 cone-beam computed tomography (CBCT) images from 117 subjects across two institutions.
• Manually segmented mandibular condyles to create ground truth datasets.
• Employed basic and cascaded 3D U-Net models for semantic segmentation.
• Conducted a stress test varying training dataset sizes to assess performance.
• Evaluated accuracy using Dice Similarity Coefficients (DSC) and Hausdorff Distance (HD).

Main Results:

• Both basic and cascaded 3D U-Net models achieved high segmentation accuracy.
• The cascaded 3D U-Net consistently outperformed the basic model.
• Segmentation accuracy, measured by DSC and HD, improved with increasing training data size.
• The largest dataset (Stage V) yielded the highest DSC values for both models.

Conclusions:

• Fully automated segmentation of mandibular condyles is feasible using 3D U-Net algorithms.
• Segmentation accuracy is directly correlated with the volume of training data.
• The developed algorithm shows potential for improving efficiency and consistency in clinical practice.