Deep Learning Approach for Objective Differentiation of Kidney Deficiency Syndrome in Reproductive Age Females: A Tongue-Face Fusion Model

Kaiwei Li^1,2Zehong Qiu³Jialing Li^1,2Feilin Deng³Kun Zou³Yihua Xu¹Chen Huang^1,2Ran Wang^1,2Zhaoji Yu^1,2Yuzhi Chen^1,2Yingxuan Zhang^4,5Zhuoliang Liu^4,5Si Chen^4,5Zhenning Su^4,5Xiaojing Liu^4,5Haiwang Wu^4,5Xiaozhen Wu^4,5Lilin Yang^4,5Yanxi Huang^4,5Songping Luo^4,5Wu Zhou^3*Jie Gao^4,5,6*

• ¹The First Clinical Medical School of Guangzhou University of Chinese Medicine, Guangzhou, China
• ²Lingnan Medical Research Center, Guangzhou University of Chinese Medicine, Guangzhou, China
• ³School of Medical Information Engineering, Guangzhou University of Chinese Medicine, Guangzhou, China
• ⁴The First Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, China
• ⁵Guangdong Clinical Research Academy of Chinese Medicine, Guangzhou, China
• ⁶State Key Laboratory of Traditional Chinese Medicine Syndrome, Guangzhou, China

Background: Kidney deficiency syndrome (KDS) is the predominant syndrome associated with gynecological reproductive system diseases in traditional Chinese medicine (TCM). However, the diagnostic method is influenced by the subjective experience of doctors, which leads to the ambiguity in differentiation of KDS and poor effect for corresponding treatment. Objective: To explore an objective syndrome differentiation method for KDS in females of reproductive age through machine learning technique. Methods: We proposed a new deep learning method for the objective differentiation of KDS in females of reproductive age. First, we simultaneously acquired 376 pairs of tongue and facial images. We divided them into a Kidney deficiency syndrome (KDS, n = 182) group and a Non-Kidney deficiency syndrome (NKDS, n = 194) group. Then, we employed two parallel DenseNet structures to extract deep features from tongue and facial images. We further used a deep supervised network strategy to better stabilize the fusion of the two deep features. We used 5-fold cross-validation to evaluate the performance by six indicators: accuracy, precision, recall, F1 score, receiver operating characteristic (ROC) and area under the curve (AUC). Finally, external validation was conducted on an independent test set consisting of 130 patients with a 1:1 ratio of KDS to NKDS cases. Results: The model based on tongue images, facial images, and the tongue-face fusion model achieved AUCs of 71.45% ± 6.39%, 89.60% ± 3.33%, and 92.08% ± 4.51%, respectively, with the highest value observed in the fusion model. In external validation, the tongue-face fusion model attained an AUC of 83.53%. Conclusions: The deep learning network model with tongue-face fusion can effectively differentiate KDS.