Research And Application Of Automatic Generation Of Medical Image Report Based On Feature Enhancement

In clinical diagnosis,radiologists write imaging reports for other doctors to choose the best diagnosis and treatment.However,the workload of doctors is heavy,which easily leads to missed diagnosis and misdiagnosis,and delays the treatment of patients.The medical report generation method based on deep learning can significantly reduce the burden on doctors and improve clinical efficiency.Medical image reports contain technical terms and long paragraphs,and due to the similar structure of medical images,the visual features of abnormal areas are not easy to capture.Existing methods focus on generating text descriptions for lesion regions,but experiments show that the differences in visual features between medical images and the semantic features of medical reports are underutilized.Based on this,this thesis conducts research on the automatic generation model of medical imaging reports based on feature enhancement,and designs and develops a medical imaging report generation system.Its main work and contributions are summarized as follows:(1)In order to capture the features of abnormal regions in input images and generate accurate text descriptions for them,a method for automatic generation of medical imaging reports based on multi-level feature differences is proposed.When calculating the feature difference between the input image and the normal image,it not only pays attention to the difference of high-level visual features,but also considers the difference of low-level visual features,so that the model can focus more on the visual features of abnormal areas.In the process of word generation,in order to prevent the model from treating the generation of visual words and non-visual words equally,an adaptive attention module is proposed to improve the accuracy of report generation.Experimental results on IU-Xray and MIMIC-CXR-JPG datasets demonstrate the effectiveness of the proposed method.(2)In order to avoid the above method of generating reports only learning from the word level and ignoring high-level semantic information,an automatic generation method of medical imaging reports based on multimodal data features is proposed.Based on the report generation branch,an image-text matching branch is designed.By coupling the two branches with each other,visual features and text features are combined to improve the accuracy of report generation.In addition,in order to alleviate the data bias in the data set and make the model tend to generate reasonable but general reports that lack abnormal descriptions,a progressive learning framework is proposed,which enables the model to learn from simple samples and iteratively learn difficult samples.On the IU-Xray and MIMIC-CXR-JPG datasets,the proposed method has improved all the evaluation indicators compared with the above method,and is better than the current mainstream method,which proves the effectiveness of the proposed method.(3)Based on the improved model,an automatic generating system for medical imaging reports was developed.The website is developed based on the VUE framework,the business logic is implemented using Spring Boot,and the deep learning model deployed on the server is accessed through the URL address.The system implements some commonly used functions and management modules.The system can help doctors automatically write medical imaging reports and ease the workload of doctors.In addition,it is also convenient for doctors to manage the health of patients.