摘要
中医舌诊可以通过舌面裂纹的测量了解身体的健康状况。考虑到采集图片分辨率不同、质量差异较大、背景复杂且不同患者舌头颜色纹理不同等因素,传统分割算法难以得到较好的结果。提出一种基于深度学习的、面向小样本数据集的舌裂分割算法。分割时先分割舌面,再进行舌面裂纹分割,以规避较为复杂的背景影响。使用迁移学习的方法将路面裂纹训练模型迁移至舌面裂纹;改进U-Net网络结构,在传递拼接路径增加SE模块以提高分割的准确性和鲁棒性;训练时使用Focal Loss(焦点损失)函数作为损失函数,使网络更加关注小目标的分割,从而提升整体分割效果。实验表明,本文方法具有更强的鲁棒性,能够应对背景、光照、纹理等图片质量问题,其MPA(平均像素精度)能够达到71.06%,MIoU(均交并比)能够达到67.35%,在视觉效果上有更好的表现。实验和对比结果也表明,该分割方法具有较高的准确性和稳定性。
Traditional Chinese medicine(TCM)tongue diagnosis can understand the health of the body through the measurement of tongue cleft.Taking into account the different resolutions,quality,color,texture and background,it is difficult for traditional segmentation methods to segment the tongue cleft to obtain better results.A segmentation method based on deep learning is proposed for the segmentation of tongue cleft under small sample data set.The tongue surface is segmented first,and then the tongue cleft segmentation can be performed to avoid complicated background effects.The transfer learning is used to transfer road crack training model to tongue cleft.The SE module is added to the transfer and splicing path of U-Net to improve the segmentation accuracy and robustness.Focal Loss function is used as the loss function during training to make the network pay more attention to the segmentation of small targets,so as to improve the overall segmentation effect.Experiments show that the method in this paper is more robust and can deal with image quality issues such as background,lighting,and texture.Its MPA(Mean pixel accuracy)can reach 71.06%,and MIoU(Mean intersection over union)can reach 67.35%,which has better performance in visual effects.Experiments and comparison results also show that the segmentation method has high accuracy and stability.
作者
王一丁
孙常浩
崔家礼
武小荣
秦雨欣
Wang Yiding;Sun Changhao;Cui Jiali;Wu Xiaorong;Qin Yuxin(College of Information,North China University of Technology,Beijing 100144,China;Beijing Bayes Health Technology Co.,Ltd,Beijing 100095,China)
出处
《世界科学技术-中医药现代化》
CSCD
北大核心
2021年第9期3065-3073,共9页
Modernization of Traditional Chinese Medicine and Materia Medica-World Science and Technology
基金
北京市科学技术委员会科技计划课题(Z181100009218012):基于微阵列镜头的老年人健康检测与分析系统研发,负责人:王一丁
