Expert-guided StyleGAN2 image generation elevates AI diagnostic accuracy for maxillary sinus lesions

Area of Science:

• Artificial intelligence in dental medicine
• Medical image synthesis
• AI-driven diagnostics

Background:

• AI research in dentistry faces data acquisition and distribution challenges.
• Developing AI tools for maxillary sinus lesions (MSL) is hindered by these data limitations.
• Existing generative models lack control over realism, diversity, and specificity.

Purpose of the Study:

• To establish an expert-guided framework to overcome data limitations in MSL AI development.
• To improve AI-based diagnostic accuracy for mucosal thickening and polypoid lesions.
• To enhance AI-assisted preoperative assessment for maxillary sinus lift procedures.

Main Methods:

• Developed a StyleGAN2 framework for expert-controlled generation of clinically relevant MSL images.
• Integrated synthetic MSL images into training datasets.
• Evaluated the impact of synthetic data on ResNet50 diagnostic performance.

Main Results:

• Generated MSL images demonstrated high fidelity (SSIM > 0.996, MMD < 0.032) and clinical validation.
• Synthetic data integration significantly improved diagnostic accuracy on internal and external test sets.
• Area under the precision-recall curve (AUPRC) increased by ~8% and ~14% for mucosal thickening and polypoid lesions, respectively.

Conclusions:

• StyleGAN2 effectively synthesized high-quality MSL images, addressing data scarcity and imbalance.
• The framework boosted diagnostic model performance for MSL detection.
• This work provides a methodological framework for overcoming data limitations in medical image analysis.