The Algorithm Research On X-Ray Image Classification And Medical Report Automatic Generation

X-ray examination is one of the commonly used means of medical imaging examination,because of its economic and rapid advantage is often used in the diagnosis of various chest diseases.In the process of clinical diagnosis and treatment,X-ray examination frequency is high,the doctor manual reading time and effort;The image similarity of different diseases is high,which is easy to misdiagnose and miss diagnosis.With the development of artificial intelligence technology,using deep learning method to realize the research of chest disease auxiliary diagnosis algorithm based on X-ray can reduce the workload of clinicians,improve the diagnostic efficiency and accuracy,and shorten the waiting time of patients,which has important research significance and clinical practical application value.Most of the current deep learning assisted diagnosis algorithms have some problems,such as inadequate feature extraction ability,not considering the correlation between different spatial features,incomplete semantic feature fusion and low accuracy of report generation algorithm.Aiming at the problems of high similarity of X-ray pathological images,insufficient ability of model feature extraction and low correlation among semantic information features,this paper focuses on the study of intelligent classification algorithm of X-ray images and automatic generation algorithm of medical reports.Aiming at the problems of high similarity of X-ray pathological images,insufficient ability of model feature extraction and low correlation between semantic information features,this paper mainly studies the X-ray image classification algorithm and the automatic generation of chest medical reports.(1)Research on intelligent X-ray Classification Algorithm based on multi-space attention.In order to improve the diagnostic quality of computer-aided X-ray images,a Multi-space attention network(MSA-Net)based on multi-space feature extraction and jump connection was proposed to solve the problems of high similarity of X-ray images and insufficient feature extraction ability.The image input features were mapped by multi-layer small convolution group,and multi-dimensional feature extraction and fusion were realized by skip join feature addition operation,so as to extract more fine-grained features of X-ray film.In the process of feature extraction,spatial attention and channel attention are used to retain the correlation information between the feature space and channel,and the weighted processing is carried out,which effectively improves the accuracy of intelligent classification of X-ray images.An experimental study on Chest X-ray Images,a public data set of chest X-rays,showed that the algorithm could reach 97.21% accuracy in binary classification(normal and pneumonia)and86.24% accuracy in tri-classification(normal,bacterial and viral pneumonia),showing outstanding classification and recognition performance.The effectiveness of the model structure is proved by the ablation experiments of the proposed module.(2)Research on Automatic medical report generation Algorithm based on MSA-Net and Shared Attention.Aiming at the problems of time-consuming and laborious writing medical reports manually by doctors,incomplete semantic feature fusion and low accuracy of automatic generation algorithm of medical reports,in order to relieve clinical pressure,assist doctors in diagnosis and treatment,and improve diagnostic efficiency,this paper,based on the encodedecoder architecture,An Attention Multitasking learning network(AML-Net)model was proposed based on MSA-Net,common attention mechanism and LSTM.The model can be divided into three parts: Firstly,the multi-dimensional feature extraction of X-ray images is further optimized by MSA-Net to effectively extract local and global features of image features,and the label prediction is carried out by MLC(multi-label classification).Then,embedded fusion of medical report text and prediction label was carried out.A parallel hierarchical common attention mechanism was used to focus on similar and similar semantic features in images and texts through similarity matrix,and weighted processing was carried out,which effectively improved the accuracy of report generation.Finally,the report generates the network classification LSTM(sentence LSTM,word LSTM),and generates the medical report word by word.In this paper,comparison and ablation experiments were conducted on IU X-Ray chest radiography report data set.The experimental results showed that AML-Net achieved good improvement in several evaluation indexes of natural language processing,among which BLUE-1 reached 0.466,with good performance of automatic generation of medical reports.In this paper,MSA-Net algorithm and AML-Net architecture are designed and proposed respectively in the research of X-ray image classification and chest medical report automatic generation algorithm,which has obtained high diagnostic accuracy.The two algorithms proposed in this paper have potential clinical application value.It is expected that they can not only help clinicians improve work efficiency,but also provide referential auxiliary diagnostic methods for the clinical screening and diagnosis of X-ray images of other diseases.