Improving machine learning-based bitewing segmentation with synthetic data.

Area of Science:

• Medical image analysis
• Machine learning in dentistry
• Artificial intelligence in radiology

Background:

• Class imbalance presents a significant challenge in medical image analysis, particularly for segmenting underrepresented structures like dental implants in bitewing radiographs.
• The pixel-level representation of implants in training and testing datasets is extremely low (0.03% and 0.07%), necessitating advanced techniques to improve model performance.

Purpose of the Study:

• To investigate the efficacy of using synthetic data generated by diffusion models and generative adversarial networks (pix2pix) to address class imbalance in dental implant segmentation.
• To compare the performance of a U-Net segmentation model trained on various data strategies, including original, synthetic, oversampled, and fine-tuned datasets.

Main Methods:

• Generation of a synthetic dataset enriched with dental implants using diffusion and pix2pix models.
• Training a U-Net segmentation model on four distinct datasets: original, synthetic, synthetic followed by fine-tuning on original, and naively oversampled data.
• Evaluation of model performance using metrics such as precision, Dice score, recall, F1 score, and ROC AUC.

Main Results:

• The U-Net model trained solely on the original dataset failed to segment implants.
• Naïve oversampling yielded the highest precision, while training exclusively on synthetic data resulted in poorer performance across all metrics.
• Fine-tuning the model pre-trained on synthetic data with the original dataset achieved the highest Dice score, recall, F1 score, and ROC AUC, outperforming other methods.

Conclusions:

• Synthetic data alone can degrade segmentation model performance, especially for minority classes.
• Fine-tuning models pre-trained on synthetic data with original data significantly improves performance for underrepresented classes, offering a promising approach for AI-driven dental imaging.