Research On Attention Convolution Networks For Intelligent Tongue Diagnosis

During thousands of years of medical practice,Chinese medicine has gradually developed and formed a medical system,which has played a great role in improving human immunity and enhancing physical health.In Chinese medicine,the tongue diagnosis occupies a very important position which can infer the human condition based on the visual characteristics of the tongue.The field of artificial intelligence has been booming in recent years,and deep learning methods based on neural networks have demonstrated capabilities beyond human doctors in certain medical imaging fields.Using deep neural network to implement intelligent TCM tongue diagnosis can help improve the effect of intelligent medicine and better assist physicians in diagnosis.This article aims to improve the performance of intelligent tongue diagnosis and improve the interpretability of intelligent tongue diagnosis.This paper(1)proposes a convolutional neural network structure calls Competitive Squeeze and Excitation(Competitive-SE)structure,which allows the attention mechanism to be more reasonably applied to the residual convolutional network,thereby improving the ability of neural networks to model images;(2)proposes a training method calls Stochastic Region Pooling(SRP)which can enhances the detail features by promoting the feature maps to respond to rich and diverse features;(3)based on the Competitive-SE and SRP methods,proposes a tongue-based disease-nature inference method,which mimics the diagnostic way of physicians and uses tongue attribute learning techniques to improve the prediction effect.Moreover,this method can provide the diagnostic process thereby increasing the interpretability of intelligent tongue diagnosis and promoting the acceptance of intelligent tongue diagnosis technology by physicians.Experimental results on image recognition datasets including CIAFR-10/100,Image Net and three Fine-grained datasets(CUB-200-2011,Stanford Cars and Stanford Dogs)show the effectiveness of Competitive-SE and SRP methods.Detailed experiments on tongue diseasenature dataset verify the effectiveness of all methods we proposed.The network that combines all methods we proposed can improve 1.41% of mean Average Precision,6.04% of Macroaveraging,and 3.91% of Micro-averaging on the tongue diagnosis task.