Tongue Diagnosis Approach To Health Identification In Traditional Chinese Medicine

Tongue diagnosis is an important part among “the four diagnostic methods” in Traditional Chinese Medicine. It’s painless, noninvasive and simplistic, which are the trends of future diagnostic methods. However, it’s also subjective and empirical, and hence limited for further development. Computerized tongue diagnosis(CTD) based on image processing and machine learning techniques has been an effective way for objectifying this process. Although there has been intensive research on this topic in past decades, many problems remain open, such as tongue body segmentation(TBS) and health identification(HI). In this thesis, we study new solutions for these two problems.TBS is a very important and difficult preprocessing step in CTD. To counter the drawbacks of requiring prior knowledge of tongue location and lacking global optimality in existing methods, a tongue localization method based on adaptive thresholding via hue histogram regression and a region-based graph cut segmentation method are proposed, and the effectiveness is experimentally demonstrated.HI is one of the objectives of CTD, for which extracting discriminating(healthy vs.diseased) features is an essential part. By taking intuition and robustness into consideration, a new feature extraction method based on color distribution homogeneity(CDH)is proposed. Experiments show that CDH features significantly outperform other state of the arts. By applying sparse principal component analysis, we further find that the color distribution similarities between tongue root and other parts convey the most important diagnostic information. In addition, statistical analysis is conducted to indicate the importance of “confounding factor”, such as age and sex, which is commonly neglected in most HI studies; resolutions to this issue based on t test are also proposed. In the last, we study anomaly detection via kernel density estimation of tongue colors as a novel formulation for HI. Different from traditional binary classification formulation, it can model the HI problem more naturally, and can perfectly handle imbalanced dataset.The effectiveness of this approach is demonstrated on a dataset from college students.