Research On Automatic Generation Method Of X-ray Image Inspection Report

The research work in this paper is to use X-ray images to automatically generate patient inspection reports.Using images to automatically generate inspection reports is still a challenging task,because this type of research requires methods to bridge the semantic gap between image information and natural language information.Meanwhile,it is even more difficult in medical scene.Because medical images are more abstract than conventional pictures.Moreover,radiology reports contain diagnostic information,more accuracy is required.In this research work,this paper proposes a neural network model based on generative adversarial networks,which can directly generate corresponding radiology report text from X-ray images.This research adopts the method of adversarial training.The generator model uses an existing model based on the Encoder-Decoder architecture to generate radiology report text from X-ray images.The Encoder is used to extract the X-ray image features,and the Decoder uses the image features to automatically generate a radiology report.This paper attempts to use the method of adversarial training to break through the performance bottleneck caused by the traditional use of cross-entropy loss training for this type of neural network model.For this reason,this paper proposes a discriminator model for adversarial training.The purpose of the discriminator is to discriminate whether the report is automatically generated by the generator model or manually written,thereby indirectly enhancing the effect of the generator model.Since the generated radiology report belongs to discrete natural language label data,it is impossible to directly update the parameters of the generator model through the reverse transfer loss gradient of the discrimination result of the discriminator model.Therefore,in the adversarial training,this paper introduce a reinforcement learning method called Self-critical Sequence Training(SCST)to update the parameters of the generator model.Finally,this paper trained and verified the model on the public X-ray inspection report data set IU X-ray.In the end,the various indicators on the test set achieved very good results,reaching 0.19 on the METEOR and above 0.4 on the ROUGE-L.And this paper also conducted a lot of comparative experiments to verify the effectiveness of generative countermeasures.The comparative experiment results show that the generative adversarial training method is superior to other methods.