Enhancing diagnostic quality in dental bitewings using transformer and GAN-Based image restoration

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Bitewing radiographs are crucial for detecting interproximal caries and assessing alveolar bone levels.
• Image compression and sensor limitations can degrade radiograph quality, impacting diagnostic value.

Purpose of the Study:

• To evaluate the effectiveness of two deep learning methods, Real-ESRGAN and SwinIR, in reconstructing details from degraded bitewing radiographs.
• To compare the diagnostic utility of images restored by these advanced AI models.

Main Methods:

• A dataset of 4,004 high-quality bitewing radiographs was used, creating paired low- and high-resolution images.
• Real-ESRGAN and SwinIR models were fine-tuned and applied to reconstruct degraded images.
• Image quality was assessed quantitatively (PSNR, SSIM) and qualitatively by a panel of dentists.

Main Results:

• SwinIR achieved superior quantitative scores (PSNR: 35.56 dB, SSIM: 0.9287) compared to Real-ESRGAN (PSNR: 31.93 dB, SSIM: 0.8227).
• Dentists preferred SwinIR for preserving diagnostic structures like tooth margins and trabecular patterns.
• Real-ESRGAN was noted for perceptual realism but less for diagnostic detail preservation.

Conclusions:

• The SwinIR model demonstrated superior performance in preserving clinically significant structures in bitewing radiographs.
• SwinIR is the preferred deep learning method for enhancing bitewing radiographs when diagnostic precision is paramount.