A Disease Auxiliary Diagnosis Method Based Onmachine Learning

The theory of Traditional Chinese Medicine(TCM)has a history of thousands of years,and the accumulated experience is a valuable asset in the medical field.The "four diagnostics" of traditional Chinese medicine are divided into Inspection,smell,inquiry,and pulse diagnosis.The diagnosis method is easy to implement,but the diagnosis results mainly rely on the subjective judgment of doctors,and the disease cannot be quantitatively described,which seriously hinders the development of traditional Chinese medicine.The study takes the objective diagnosis of TCM as the research focus,establish the database of face,tongue,hand and pulse,and provides theoretical basis for the diagnosis of TCM by using the neural network algorithm.On the basis of summarizing previous research,a classification method of TCM symptoms on a small data set was proposed,which mainly includes the following work:1.Establish a dataset of four diagnosis of TCM.The objective diagnosis of TCM is a major trend in current research,but there are very few TCM datasets publicly available.The study takes the syndrome of "liver depression and spleen deficiency",one of the three major chronic diseases,as the research object,using four diagnostic instruments of traditional Chinese medicine and photoelectric pulse acquisition device to collect the dataset of face,hand,tongue and pulse.After image preprocessing,the four diagnostic data set of traditional Chinese medicine is established.For the tongue diagnosis image,according to the characteristics of H component in HSV space,the tongue coating part is segmented from the background image.Combined with image morphology method and contour area comparison method,the tongue can be segmented better.In the pulse signal processing,the signal is denoised and the baseline drift is removed,and a feature extraction method based on waveform segmentation is proposed.According to the trough of the pulse waveform,the period is divided,and the one-dimensional time series signal is converted into a two-dimensional image.After normalizing the data length and the data value,the one-dimensional pulse signal is converted into a grayscale image.2.In view of the problem of small sample size,online and offline data augmentation is performed on small sample data,using five enhancement methods of rotation,shearing,zooming,horizontal flipping,and translation and their combinations.By comparing the changes in the accuracy of the model before and after data augmentation,the most accurate data augmentation method on each dataset was obtained.3.Classification and identification of four diagnostic datasets.This study uses the Inception-V3 network to classify and recognize four kinds of traditional Chinese medicine images,and explores the effect of training epochs on the classification accuracy.In the course of the experiment,aiming at the problem that the samples do not obey the same distribution,the data shuffling was carried out,and the accuracy of the data shuffling was proved effective by comparing the accuracy before and after the shuffling.The diagnostic accuracy rates of the face diagnosis,hand diagnosis,tongue diagnosis,and pulse diagnosis models are 77.0%,92.9%,80.8%,and 72.2%.The experimental results show that the neural network model has achieved good classification on small sample data sets.4.In order to further improve the recognition accuracy of the model,three methods of fourdiagnosis were compared,which are multi-feature fusion method,multi-model fusion method and the Dempster-Shafer theory.The multi-feature fusion method splices the four-diagnosis images of patients and sends them to the neural network for classification training.The multi-model fusion method uses the result majority vote and the result weighted average method to predict the samples.The Dempster-Shafer theory fuses the models in pairs.The experimental results show that the Dempster-Shafer theory can achieve the same prediction accuracy as the traditional multi-model fusion method on the face,hand,and tongue data sets with less fusion models,which shows the best model fusion performance.