Evaluating masked self-supervised learning frameworks for 3D dental model segmentation tasks.

Area of Science:

• 3D computer vision
• Medical imaging analysis
• Artificial intelligence in dentistry

Background:

• Automated computer-aided treatment planning in dentistry relies on deep learning with 3D dental models.
• High-accuracy model development is hindered by the scarcity of labeled medical data.
• Masked self-supervised learning (SSL) offers a potential solution to data scarcity challenges.

Purpose of the Study:

• To investigate the effectiveness of four masked self-supervised learning frameworks (Point-BERT, Point-MAE, Point-GPT, Point-M2AE) for 3D dental models.
• To evaluate the impact of pre-training on downstream tasks like tooth and brace segmentation.
• To determine the optimal conditions for applying masked SSL in dental applications.

Main Methods:

• Pre-training of four masked SSL frameworks on over 4000 unlabeled 3D dental models.
• Fine-tuning the pre-trained models on the Teeth3DS dataset for tooth segmentation.
• Fine-tuning on a custom dataset for braces segmentation.
• Experimental evaluation of performance enhancement in downstream tasks.

Main Results:

• Pre-training significantly enhances the performance of downstream tasks, particularly with limited or imbalanced labeled data.
• The benefits of masked SSL pre-training are most pronounced under data scarcity conditions.
• Performance gains diminish as the amount of available labeled data increases.

Conclusions:

• Masked self-supervised learning is a viable strategy to improve deep learning models for dental applications, especially in data-limited clinical settings.
• Pre-training with masked SSL can improve the accuracy and clinical usability of automated treatment planning systems.
• Understanding the relationship between labeled data availability and performance gains is crucial for effective implementation.