Research On Obstetrics And Gynecology Ultrasound Diagnosis Support Method Based On Deep Learning

The rapid growth of the Internet has brought about explosive growth in medical data.Compared to human capabilities,computers are better suited to handle huge and complex data and extract the information within it.To address the challenge of the information explosion,there is an urgent need for technicians to use computer tools to solve efficiency problems in the diagnosis and treatment process and to relieve the pressure on medical practitioners.Traditionally,medical diagnosis requires a great deal of expertise and experience from doctors.As the number of patients increases rapidly,this leads to significant labor costs and time overhead in the diagnosis process.With the development of artificial intelligence in the field of wise information technology of med,the use of deep learning technology for end-to-end medical diagnosis support with the content to be diagnosed as input and the diagnosis result as output can effectively save labor costs and improve the efficiency of diagnosis and treatment,which is of great significance in clinical application.Based on the actual needs of a Grade-A tertiary care hospital,this paper focuses on a text-based diagnosis support task for obstetrics and gynecology ultrasound.There has been less research on related issues in existing work,mainly due to the following challenges for this type of task.From a data perspective,there are few publicly available datasets in Chinese,and research for Chinese medical texts has started late.In addition,the content of medical texts is more specialized than in the general domain,and the lack of obvious word separation markers makes Chinese medical texts more difficult to handle.From a task perspective,there is little research on similar downstream tasks for text diagnosis.Therefore,it is an important question for this study to choose an applicable solution and exploit the causal and sequential relationships between medical texts,and ultimately propose a suitable task evaluation scheme and an effective diagnosis support method in conjunction with the characteristics of the diagnosis task.To address the above problems and challenges,this paper proposes a hybrid diagnosis support method for OB/GYN ultrasound based on deep learning,with the aim of automatically generating the corresponding diagnostic results based on the ultrasound description text in the report.In addition,this paper provides a usable Chinese dataset of OB/GYN ultrasound reports.The specific work is as follows.1.By collecting real data from the ultrasound department of a Grade-A tertiary hospital after removing patients’personal sensitive information,this paper builds a publicly available dataset of OB/GYN ultrasound reports,which contains 180,000 complete ultrasound reports and nearly 30,000 annotated data.Based on this dataset,this paper proposes an unsupervised OGBD algorithm for OB/GYN ultrasound dictionary construction to alleviate the problem of difficulties in processing medical texts.2.Based on the data and task characteristics,this paper proposes a hybrid diagnosis support method for OB/GYN ultrasound based on a combination of label classification and text generation.The label classification approach converts the diagnostic results into a sequence of labels through manual annotation,while the text generation approach does not modify the form of the diagnostic results.For each of these two approaches,an evaluation scheme is proposed to mitigate the misclassification problem.The alternative models available are then proposed according to the task characteristics,and the best-performing model is selected based on this scheme.Finally,the paper proposes an optimization algorithm based on relational extraction,OGPM,which exploits the causal relationships between texts to improve the overall performance of the method.3.Based on the above algorithm and dataset,this paper designs and implements a prototype system for this diagnosis support method and a visual interface that is easy to use by doctors and patients.In addition,the performance of the method is validated and the prediction results are subdivided into 10 components to analyze the effectiveness of the method individually.The experimental results show that the diagnosis support method proposed in this paper achieves good accuracy,validity,and usability.The label classification method can finally achieve an accuracy of 92.41%,the precision of 94.59%,recall of 95.15%,F₁-score of 94.23%,and F₁-micro of 93.61%.The best results for the text-generative approach were 0.9437 for Rouge-1,0.9201 for Rouge-2,0.9418 for Rouge-L,and 0.8844 for BLEU.7 out of the 10components of the report achieved 0.85 or higher for the four metrics mentioned above.The diagnosis support method is somewhat generalizable and can be transferred to similar studies in other medical fields.If the method is combined with automatic ultrasound image recognition methods,it will be possible to achieve fully automated ultrasound examinations in the future.