A noise adaptive deep learning model for dental caries analysis in low-dose X-ray images

Dental caries remains one of the most common diseases globally; therefore, early and reliable detection is essential for reducing disease progression and treatment costs. However, existing state-of-the-art algorithms for low-dose dental images are still suffering from Poisson noise, which affects pixel values and edge information, a lack of well-annotated datasets, and limited interpretability. To solve these issues, we propose a novel adaptive deep learning model for dental caries analysis in noisy dental images. Initially, a novel layer discrimination Poisson noise reduction algorithm is proposed to reduce Poisson noise from low-dose dental X-ray images. Our proposed noise reduction algorithm improves dental structures by detecting only the noisy pixels without affecting the uncorrupted pixels. It groups and structures corrupted and uncorrupted pixels into separate layers and denoises and restores only the corrupted pixels. Following that, an upgraded lightweight MobileNetV3 is applied to extract meaningful details from the enhanced X-ray images. Finally, a swin transformer, coupled with a U-Net-inspired decoder, is applied to detect caries and generate accurate masks that show pixel-level detail in enhanced X-ray images. Our model was evaluated using multiple datasets and achieved an accuracy of 94.11%, a sensitivity of 96.03%, an AUC of 0.98, and a classification accuracy of 96.5%. The results demonstrate the effectiveness of our proposed Poisson noise reduction algorithm in achieving high PSNR and low MSE. Our proposed model achieves high accuracy, sensitivity, specificity, F1 Score, AUC, Kappa, and MCC, advancing the automated detection of dental caries from low-dose dental X-ray imaging.