Automatic and visualized grading of dental caries using deep learning on panoramic radiographs

Caries grading plays a significant role for oral health management and treatment planning. Grading caries on panoramic image is a challenging task due to complication and diversity of gray distribution. In this paper, we proposed an automatic and visualized caries grading method for panoramic image using deep learning-based tooth anatomical segmentation and regions intersection judgment to achieve a consistent grading process with dentist. To achieve accurate semantic segmentation, a modified U-Net model by adding ASPP module and boundary loss is applied to segment caries, enamel, dentin, and pulp tissue region. Then a visualized process is conducted to judge the intersection of carious region and decision-making line for grading of shallow, medium, deep caries. Experimental results demonstrate our method achieves promising grading performance. Moreover, we validated that our proposed two-stage caries grading method outperform deep learning classification models. Ablation analysis of anatomical segmentation performance was also investigated, and the compared results show that our proposed modified U-Net model can obtain more accurate region and boundary to improve grading results. Some mis-graded cases were finally detailed analyzed. Our proposed caries grading approach has great potential for clinical aided diagnosis and automatic chart filling on panoramic radiographs.