Research Of Medical Data Analysis Based On Information Fusion

With the development of informatization in medical institutions,mass heterogeneous multi-source medical data has been generated during multiple proceedings of medical services like diagnosis,therapy,and follow-up.It is remarkable that privacy concerns lead to data island issues,which brings challenges and obstructions to data-driven research.Moreover,it is difficult to utilize and analyze medical data efficiently and comprehensively for certain medical services like diagnosis.In response to the aforementioned problems,explore the technique of federated learning and reinforcement learning to solve the problem of data island and medical information fusion issues respectively.Taking thyroid cancer diagnosis as an example,research the feasibility of the medical data analysis system.Thyroid disease is the most pervasive disease in the endocrine system,the diagnosis of which includes several ways like ultrasound,blood examination,and pathology.Precious diagnosis in early-stage helps reduce unnecessary fine-needle aspiration biopsy and relieve patients’ pain.First,training deep neural networks between medical institutions by collaboratively passing parameters,in order to classify the thyroid ultrasound image.Considering privacy issues,add adequate disturbance in the process of parameters transferring to avoid data leakage caused by parameter leakage.Then,design a reinforcement learning framework,which can fuse various diagnostic information at the output level,thus promote metrics like accuracy,sensitivity,and specificity.Besides,with the development of techniques of information extraction towards different information sources,the scalable framework is capable of containing more complex diagnostic information.Based on the collected clinical data,experiments verify the feasibility of federated learning,showing that there is no obvious loss in key metrics with the guarantee of privacy.Compare to the baseline methods,the reinforcement learning model can rise about 2-9% in accuracy,2-5% in specificity,and 4-7% in sensitivity,which means it can facilitate aided diagnosis service and owns the capability of information fusion in sense.