基于YOLOv8改进的牙齿X光医学图像分割模型

Dental X-Ray Medical Image Segmentation Model Improved Based on YOLOv8

摘要: 医学牙齿图像对于帮助医生根据图像中的器官、组织和病变视觉表现，来快速进行诊断和临床干预。其中牙齿分割和编号是口腔分析诊断的重要开始。在牙科治疗中，X光片是一种常见的诊断工具，在诊断中被医生用于检查牙齿、牙龈、颚骨和口腔骨骼结构的状态。自动准确地分割牙齿位置和区域，对牙齿分割模型非常重要，由于牙科X射线图像的对比度差，特征提取困难，导致预测掩膜质量较低。本文为了提高牙齿X光图像分割的质量提出了YDBD模型。该模型基于YOLOv8改进的EfficientNet骨干网络中加入了空间注意力机制，增加了牙齿位置的识别的准确性。在其Neck结构中，使用了动态上采样来减少边缘伪影，并加入高层特征图拼接，对边缘特征进行更好的提取，提高预测掩膜贴合度。实验结果表明准确性和平均精度均值方面对比原模型有部分提升。

Abstract: Medical dental images play a crucial role in helping doctors make rapid diagnoses and clinical interventions based on the visual manifestations of organs, tissues, and lesions in the images. Among them, tooth segmentation and numbering are important starting points for oral analysis and diagnosis. In dental treatment, X-rays are a common diagnostic tool, which are used by doctors to examine the condition of teeth, gums, jawbones and oral skeletal structures during diagnosis. Accurate and automatic segmentation of tooth positions and regions is of great significance for the tooth segmentation model. Due to the poor contrast of dental X-ray images and the difficulty in feature extraction, the quality of the prediction mask is relatively low. In this paper, to improve the quality of dental X-ray image segmentation, the YDBD model is proposed. This model is based on the improved EfficientNet backbone network of YOLOv8 and incorporates a spatial attention mechanism, which enhances the accuracy of tooth position recognition. During upsampling, dynamic upsampling is used to reduce edge artifacts and high-level feature maps are concatenated to better extract edge features, thereby improving the fitting degree of the prediction mask. Experimental results show that there is a partial improvement in terms of accuracy and mean average precision compared to the original model.