舌象图像深度特征在慢性失眠疗效评价中的应用

Application of deep features of tongue images in evaluating efficacy of treatments for chronic insomnia

目的基于舌象深度特征构建慢性失眠常见中医证型的疗效评价模型。方法基于220例健康人舌象,应用深度卷积神经网络(ResNet50)分别对241例痰热扰心证、185例心脾两虚证、266例心肾不交证慢性失眠患者舌象进行二分类学习,得到验证集超过95%分类准确率的3个不同基准模型,将其参数固定后,再将中药治疗前后的图像分别输入该模型,获得相应的概率输出,即该病例治疗前后的健康似然度,并进行分析。结果治疗期间,慢性失眠不同证型中药治疗有效病例舌象特征健康似然度呈线性升高趋势;中药治疗无效病例的舌象特征健康似然度的变化情况与中医证型有关,分别呈线性下降(痰热扰心证)、先升后降至治疗前水平(心脾两虚证)以及缓慢上升(心肾不交证)的趋势。结论基于舌象深度特征的慢性失眠疗效评价方法及可视化呈现,具有良好的客观性、可读性。

Objective To construct an efficacy evaluation model for common traditional Chinese medicine(TCM)syndromes of chronic insomnia based on deep features of tongue images.Methods Based on the tongue images of 220 healthy controls,the deep convolutional neural network(ResNet50)was used to conduct binary classification learning on the tongue images of 241 patients with chronic insomnia of phlegm-heat disturbing the heart syndrome,185 patients with heart-spleen deficiency syndrome,and 266 patients with heart-kidney non-interaction syndrome.Three different benchmark models with a classification accuracy of more than 95%in the verification set were obtained.After fixing their parameters,the images before and after TCM treatment were input into the model respectively to obtain the corresponding probability output,that is,the health likelihood of the case before and after treatment,and analysis was carried out.Results During the treatment period,the health likelihood of tongue image features of effective cases of TCM treatment for different syndromes of chronic insomnia showed a linear increasing trend;the change of health likelihood of tongue image features of ineffective cases of TCM treatment was related to TCM syndromes,showing a linear decreasing trend(phlegm-heat disturbing the heart syndrome),first increasing and then decreasing to the pre-treatment level(heart-spleen deficiency syndrome),and a slow increasing trend(heart-kidney non-interaction syndrome),respectively.Conclusion The efficacy evaluation method and visual presentation of treatment for chronic insomnia based on deep features of tongue images have good objectivity and readability.