This study aimed to build a home use deep learning segmentation model to identify the scope of caries lesions. A total of 494 caries photographs of molars and premolars collected via endoscopy were selected. Subsequently, these photographs were labeled by physicians and underwent segmentation training by using DeepLabv3+, and then verification and evaluation were performed. The mean accuracy was 0.993, the sensitivity was 0.661, the specificity was 0.997, the Dice coefficient was 0.685, and the intersection over union (IoU) was 0.529. Therefore, the present deep learning segmentation model can identify and segment the scope of caries.
本研究目的是建立一个可家用的能直接显示龋病范围的深度学习分割模型。收集解放军总医院第一医学中心口腔科门诊2019年9月至2021年6月共494张用内窥镜采集的、含有龋齿的磨牙和前磨牙照片，由医师进行标注后用DeepLabv3+进行分割训练，随后进行验证和评估。建立的深度学习分割模型识别龋病的平均准确度为0.993，灵敏度为0.661，特异度为0.997，Dice系数为0.685，并交比（IoU）为0.529。本研究建立的深度学习分割模型可以识别并分割出龋病范围。.