[[alternative]]Using Convolutional Neural Network for Dialectic Classification of Tongue Diagnosis in Traditional Chinese Medicine

[[abstract]]舌診在中國醫學上有著深遠的歷史，同時也是中醫最常使用來診斷的方法之一，透過辨識舌體舌苔的變化可以診斷疾病或體質。舌診的方便性在於它的簡單性和即時性。然而，傳統的舌診有它無可避免的侷限性。如舌診診斷必須仰賴於中醫專業人員的經驗和知識，而環境因素如光照、受試者因素如病患診視位置皆有可能對舌診的診斷會造成差異，而影響治療成效。先前的學者研究多半是針對舌形與舌體進行分析，如：分析齒痕舌與病人疾病臨床症狀之關聯性。很多研究指出舌頭的特徵、形狀、顏色甚至是舌苔等生理變化顯像都可以提供醫師客觀的醫療診斷依據，但也經常會因醫師個人臨床經驗而有不一樣的診斷結果。 本研究透過舌診文獻書籍掃描舌診影像及預處理，同時結合深度學習(Deep Learning)卷積神經網路的演算法進行影像辨識分析技術。為提高模型準確率，在研究中也利用預訓練模型及影像擴增的方法建立辨識模型系統並建置部署網頁平台，提供醫療人員更準確更有依據之中醫舌診輔助機制。本研究在疾病證侯上根據中醫舌診臨床圖解文獻將證侯種類分成五大類型，再運用多種評估指標檢視模型之訓練效能。研究發現若先利用訓練資料集建立訓練模型成果，再透過舌診影像資料驗證集驗證模型的訓練會可以呈現高準確度的辨識分析校能。最後，依據機器學習建置的模型本研究部署了網頁輔助診斷系統平台，提供民眾與醫療人員更便利、準確的舌診系統進行健康狀況檢視，以期達到自我健康檢測及輔助臨床中醫師診斷的效果。本研究期待能在日後擴增更多舌診影像資料並進行舌形及證侯關聯的舌診影像辨識，以提升臨床醫師於舌診判斷疾病時的準確性。[[abstract]]Tongue diagnosis has a long history in Chinese medicine, and it is also one of the most commonly used diagnostic methods in Chinese medicine. It can diagnose diseases or constitutions by identifying changes in the tongue coating. The convenience of tongue diagnosis lies in its simplicity and immediacy. However, traditional tongue diagnosis has its unavoidable limitations. For example, tongue diagnosis must rely on the experience and knowledge of traditional Chinese medicine professionals, and environmental factors such as light, subject factors such as the patient's viewing position may cause differences in diagnostic results then affect the therapeutic effect.Most of the previous research by scholars focused on the analysis of tongue shape and body, such as analyzing the relationship between teeth printed tongue and the clinical syndrome of the patient. Many studies have pointed out that the characteristics, shape, color, and even the physiological changes of tongue coating can provide doctors with an objective medical diagnosis basis, but there are often different diagnosis results due to the individual doctor’s clinical experience.This research scanned and preprocessed tongue diagnosis images from tongue diagnosis literature and books, then employed deep learning (Deep Learning) convolutional neural network (CNN) algorithm for image recognition and analysis. To improve the accuracy of the model, the pre-training model and the method of image augmentation are used to establish an identification model system. In this study, the disease syndromes were divided into five types according to the clinical graphic literature of TCM tongue diagnosis, and then various evaluation indicators were used to examine the training performance of the model. The research found that use machine learning technology and training data set to generate a recognition model, then using validation data set to validate the model can obtain a high recognition performance. The research deployed a tongue-diagnosis web platform to provide medical professionals and public individuals with a more accurate and more evidence-based auxiliary mechanism for self-health status inspection by TCM tongue diagnosis. The study expects to augment and analyzes more clinical tongue image data to explore the relationship between tongue shape and clinical syndromes and improve the accuracy of physicians in clinical diagnosis by tongue diagnosis