Integrated method for grading diagnosis of dental fluorosis combined with segmentation and classification

Endemic fluorosis is chronic fluorosis caused by excessive accumulation of fluorine in the body. The early symptom of fluoride toxicity is dental fluorosis. In severe cases of poisoning, it can lead to skeletal fluorosis. Screening for fluorosis is conducted almost yearly in remote fluorosis areas, but the lack of medical resources often leads to misdiagnosis or missed diagnosis. Therefore, applying deep learning technology to diagnose dental fluorosis is significant. Through field research and literature review, we discovered that the current screening for dental fluorosis relies entirely on doctors’ knowledge and experience without applying deep learning in relevant studies on automated diagnosis. Based on the analysis of images, we noted that the lesions display irregular shapes and indistinct borders. Additionally, the appearance of lesions can be influenced by various factors, such as extrinsic staining and illuminance. They present challenges in diagnosing dental fluorosis. So, we proposed a two-stage methodology for grading the diagnosis of dental fluorosis. In the first stage, we proposed an improved U-Net based on large kernel convolution for tooth region segmentation. Additionally, we designed a pixel-association iterative algorithm to optimize the segmentation results. In the second stage, we devised a dual-branch fusion classifier based on CNN and Transformer, which ensured accurate classification even when the lesion features occupy only a tiny proportion of the entire image. We established the first dental fluorosis dataset and evaluated our proposed method, which has achieved satisfactory results in the grading diagnosis of dental fluorosis after conducting many experiments.

Biomedical Signal Processing and Control, 96, 106510