The Research On Sublingual Vein Segmentation And Its Pathological Characteristics Analysis

In Traditional Chinese Medicine, diagnosing by inspection is a traditional method with a long history, in which there is an important class named tongue diagnosis. There are mainly two subclasses in tongue diagnosis, namely lingual diagnosis and sublingual vein diagnosis. There already have many research results in lingual diagnosis after years of research and development, but little research results in sublingual vein diagnosis, for this reason, research in sublingual vein diagnosis has significant developing potential.In this dissertation, we are particularly interested in potentially valuable features in sample sublingual images which might provide certain clues and indications in diagnosing which can be used and exploited to diagnose. Specifically in this dissertation, we are performing following procedures with intention of seeking for these valuable features in mind: sampling and pre-processing of sample images, isolation of sublingual vein part in images, feature extraction and optimization, and cluster analysis on features of sublingual veins.We are cooperating with hospitals, taking image samples of sublingual veins of certain patients, with them also corresponding biochemical indicators, diagnosis from doctors, and put all these information into tags of corresponding samples. So far we’ve already collected about 2,000 samples with tags. In this dissertation specially, we are mainly interested in sample images of patients with lung cancer, hypertension, breast cancer, kidney disease, insomnia, diabetes, gastritis, tumor and samples of healthy people. In this dissertation we do color correcting on sample images by using a method based on polynomial principle, and we proposed an interactive segmentation algorithm based on region growing in HSI and LUV color space to do image segmentation to isolate crucial part of the image from uninterested part.Based on 1000 segmented sample sublingual vein images, we built color space of sublingual vein images, and by using k-means cluster method we obtain color features of sublingual vein in that particular color space. Meanwhile, after doing sublingual vein segmentation we extracted color features on RGB and HSV color space, aquired the length, width of two sublingual veins, the ratio between length and width, and similar geometrical features. After that, we tried many combination of these values to maximize the accuracy.Finally we use selected crucial feature values to perform cluster analysis on images of sample sublingual vein, and try to distinguish samples with disease from samples without. In testing binary classification algorithm we tested many classifier algorithm by comparing their classifying results on sublingual veins samples, and eventually decided to use SVM binary classifier which has the best performance among many, which has average accuracy of 80.88%, and achieved a satisfying accuracy of 89.34% in certain samples such as type II diabetes and healthy samples. This is an indication that we are on the right track and that this method might indeed work. In multi-classification algorithm, we built SVM decision tree multiple classifiers, which has accuracy of 70.19% in samples of healthy people, samples of insomnia, and samples of breast cancer, and this is another sign of that SVM decision tree algorithm do have relatively acceptable performances in multiple classification problems, and that diagnosing by analyzing carefully taken images of sublingual vein might work in more cases, and it might, in the future, be a useful tool in diagnosing.