Deep learning-based tongue image recognition for hypertension with phlegm-dampnessconstitution(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2025年第4期
- Page:
- 534-541
- Research Field:
- 医学人工智能
- Publishing date:
Info
- Title:
- Deep learning-based tongue image recognition for hypertension with phlegm-dampnessconstitution
- Author(s):
- ZHU Qianqian1; WANG Lan1; JIANG Nan1; DONG Changwu2
- 1. College of Chinese Medicine, Anhui University of Chinese Medicine, Hefei 230012, China; 2. Eminent Physician Clinic, The SecondAffiliated Hospital of Anhui University of Chinese Medicine, Hefei 230061, China
- Keywords:
- tongue manifestation; deep learning; hypertension; phlegm-dampness constitution; neural network
- PACS:
- R318
- DOI:
- 10.3969/j.issn.1005-202X.2025.04.016
- Abstract:
- Objective To objectively identify whether people with phlegm-dampness constitution suffer from hypertension ornot using deep learning semantic segmentation model and residual neural network, so as to promote the modernization oftongue manifestation research, and provide a more objective and scientific basis for clinical decision-making in traditionalChinese medicine (TCM). Methods The tongue regions of 547 subjects were outlined and labeled using the Label Me imagelabeling tool, followed by tongue body segmentation using the U-Net segmentation algorithm which separated the tonguebody from the complex background. In the subsequent study, 3 deep learning models, namely ResNet-34, ResNet-50 andYOLOv5, were used to classify the tongue manifestations of hypertensive patients and the sub-health both with phlegmdampness, and to construct the corresponding classification models whose performances were objectively evaluated bydrawing confusion matrix and calculating F1 value and accuracy. Results The experimental results showed that all 3 modelsperformed well in the classification task. ResNet-34 vs ResNet-50 had F1 values of 91.46% vs 92.08%, accuracies of 92.87%vs 93.05%, precisions of 90.48% vs 95.26%, and recall rates of 92.89% vs 89.11%. YOLOv5 had an overall accuracy of85.6%, achieving 85.3% and 85.7% accuracies in the specific classifications for hypertensive patients with phlegm-dampnessand the sub-health with phlegm-dampness. Conclusion All 3 models (ResNet-34, ResNet-50 and YOLOv5) performed wellin the classification task, with ResNet-50 being the best. It proves that the deep learning model can better accomplish the classification and recognition of tongue manifestations, which reflects the great potential of deep learning in the automatedclassification for TCM tongue diagnosis, and also provides a strong technical support for the modernization and objectivity ofTCM diagnosis.
Last Update: 2025-04-30