Constructing a classification model for coronary artery lesion severity based on tongue manifestations and deep learning

To develop a deep learning-based intelligent tongue manifestation recognition model for the noninvasive assessment of coronary artery disease severity.

This study prospectively enrolled 147 patients with coronary heart disease who were hospitalized in the Department of Cardiology at Dongzhimen Hospital, Beijing University of Chinese Medicine. General patient information was collected, including gender, age, past history, laboratory test result, and medication use. According to the Gensini score system, the severity of coronary artery lesions was stratified into mild group (n = 44), moderate group (n = 53), and severe group (n = 50); and 147 patients were randomly divided into the training set (n = 103) and the test set (n = 44) at a ratio of 7∶3 using the random nomber table method. Standardized tongue images were acquired using the Daosheng DS01-B tongue and facial diagnosis information acquisition system. The improved UNet++ model (ISE-UNet++), incorporating multi-scale convolution and channel attention mechanisms, was employed for precise segmentation of the tongue region. A residual neural network (ResNet) deep learning model was then constructed to classify the severity of coronary lesions based on tongue images. Finally, gradient-weighted class activation mapping (Grad-CAM) was used to visualize the discriminative regions of tongue images identified by the deep learning model across different severities of coronary artery disease.

A total of 147 patients with coronary heart disease were included in this study, among whom 73 were male (49.7%) and 74 were female (50.3%), with an average age of (73.05±8.24) years. In the tongue image segmentation task, the ISE-UNet++ model outperformed the original UNet++ model, showing improvements in mean intersection over union (MIoU), mean pixel accuracy (MPA), and overall accuracy. For the classification of coronary artery disease (CAD) severity, the ResNet-50 model demonstrated the best performance on both the training and test sets. Specifically, the sensitivity was 84.8% and 74.5%, specificity was 0.829 and 0.756, precision was 0.741 and 0.788, recall was 0.805 and 0.819, F1-score was 0.790 and 0.813, accuracy was 0.809 and 0.778, Kappa coefficient was 0.777 and 0.715, and AUC was 0.880 and 0.854, respectively—all metrics surpassing those of ResNet-18 and ResNet-34 models. Grad-CAM visualization analysis revealed that in patients with mild coronary lesions, the model′s attention was focused on the tongue tip, whereas in more severe cases, the attention shifted progressively toward the middle and root regions of the tongue.

The combination of deep learning and tongue image analysis offers a novel noninvasive approach for assessing coronary artery disease severity, with promising potential for clinical application.