深度学习辅助中医望诊的抑郁症诊断模型研究OA

Objective To develop a deep learning-based facial diagnostic model for depression,providing a novel approach for early recognition of depression through Traditional Chinese Medicine(TCM)facial inspection.Methods A total of 437 participants were enrolled,including 210 patients with depression of the liver-qi stagnation and spleen-deficiency pattern,114 non-depressed individuals with the same TCM pattern,and 113 healthy controls.Facial images were collected following standardized procedures.After excluding blurred,occluded,or abnormally posed images,data augmentation was performed using rotation,scaling,flipping,translation,and Gaussian noise.The dataset was divided into training,validation,and test sets in an 8:1:1 ratio.Classification models were constructed using EfficientNet,MobileNet V3,and ResNet18.Model performance was evaluated in terms of accuracy,parameter size,convergence speed,and confusion matrix.Class activation mapping(CAM)was employed to visualize the regions of facial attention contributing to model decisions.Results(1)EfficientNet achieved the highest accuracy(98.6％),with a moderate parameter size(4.01 M)and fast convergence,followed by MobileNet V3(accuracy 92.7％,1.52 M parameters),while ResNet18 showed the lowest accuracy(92.2％)and the largest parameter size(11.18 M).(2)The confusion matrix revealed that EfficientNet exhibited the lowest misclassification rate among the three groups.(3)CAM visualization demonstrated that EfficientNet not only focused on the periocular region but also on the perioral area when making predictions.Conclusion Deep learning-based facial image analysis can effectively extract depression-related facial expression features,providing objective evidence to support intelligent TCM facial diagnosis and assist in the recognition of depressive disorders.