Application Of Artificial Intelligence In Tongue Diagnosis And Syndrome Differentiation Of Gastroesophageal Reflux Disease

Objective:The study was focused on gastroesophageal reflux disease(GERD).By collecting tongue image data of patients with gastroesophageal reflux disease,a tongue image database with syndrome type as the classification standard was established.The tongue images of dif’ferent tongue colors and tongue coating were objectively classified by cluster analysis,and the relationship between tongue image features and corresponding syndromes of GERD patients was quantitatively analyzed.The tongue diagnosis recognition model was trained by using the open platform of artificial intelligence.Methods:Camera Raw 7.0 was used to correct the color of the image and restore the real color of the tongue image in a stable lighting environment with auxiliary light source.Photoshop was used to segment the tongue image and remove the interference pixels such as lips,teeth and face.In the color space of Lab,MATLAB R201 8a was used to extract quantized tongue image data L(brightness),a(red-green),and b(yellow-blue)from different regions(tongue root,tongue center,tongue tip,left and right tongue edges).Tongue color,coating thickness and coating color were used as the classification criteria,and SPSS 19.0 was used for systematic clustering of tongue images.Using the clustering results,the tongue color,coating thickness and coating color recognition models were trained on the open platform of baidu artificial intelligence.Results:A total of 110 cases were included in this study,and 328 tongue images were collected.1.Distribution of syndrome typesThere were 14 patients with syndrome of liver-stomach heat stagnation(12.73%),5 1 patients with sy,ndrome of disharmony between liver and stomach(46.36%),20 patients with syndrome of reverse qi duo to spleen deficiency(18.18%),16 patients with syndrome of qi stagnation and phlegm obstruction(14.55%),1 patient with syndrome of blood stasis in Zhongwan(0.91%),and 8 patients with deficiency syndrome of stomach Yin(7.27%).2.Lab value characteristics of each region L(middle tongue),>,L(root,tip of tongue),L(right side of tongue),>,L(left side of tongue),a(tip of tongue),>,a(middle and side of tongue),>,a(root of tongue),b(root of tongue),>,b(middle,side of tongue,tip of tongue).3.Clustering analysis Taking the L and a values of the whole tongue,left and right tongue edges,middle tongue and tip of tongue as clustering var-iables,systematic tongue color clustering was carried out,which was divided into 5 categories.Combining with the tree diagram and expert consensus,appropriate combination was made,which could be divided into "light red tongue","red tongue" and "light white tongue".Taking the a value of tongue root and tongue as clustering variable,the systematic clustering of tongue coating thickness is divided into 5 categories,which can be divided into "thick coating","thin coating" and "little/no coating" if appropriate combination.After the "little/no moss" tongue image was removed,the systematic clustering of tongue coating color was carried out by taking the tongue root and the median b value of tongue as clustering variables,which was divided into three categories,including "white coating" and "yellow coating".The above clustering results were basically consistent with the traditional tongue diagnosis results.4.Artificial intelligence identificationThe tongue recognition model was trained based on the clustering results.The accuracy of tongue color recognition was 87.0%,accuracy 81.7%and recall 75.7%.The accuracy rate of tongue coating thickness identification was 88.8%,the accuracy rate was 92.7%,and the recall rate was 76.7%.The accuracy rate of tongue coating color recognition was 98.9%,the accuracy rate was 96.9%,and the recall rate was 99.3%.After verification,the accuracy of tongue color recognition is 90.6%,the accuracy is 83.6%,and the recall rate is 78.9%.The accuracy rate of tongue coating thickness identification was 89.8%,the accuracy rate was 91.8%,and the recall rate was 79.0%.The accuracy rate of tongue coating color recognition was 97.7%,the accuracy rate was 98.7%,and the recall rate was 93.8%.Conclusion:1.Tongue color can be well restored through stable lighting environment and color correction in later period;2.Compared with other color Spaces,Lab color space is more suitable for the objective analysis of tongue image.3.Quantitative data of different regions of tongue image were used for cluster analysis,and the obtained classification results were basically consistent with the traditional tongue diagnosis;4.According to the clustering results in the study,the artificial intelligence platform can easily and quickly identify tongue features.