Optimization Technology Of CXR Image Analysis Based On Transformer

Chest X-Ray images(CXR)have essential application value in diagnosing lungrelated diseases.At present,computer vision technology is widely used to analyze and process CXR images.However,existing research usually directly migrates computer vision algorithms that process natural images to CXR images but rarely considers problems such as model inference performance and low data utilization.This paper optimizes the CXR image analysis and processing algorithm based on Transformer to solve these problems,focusing on the two tasks of lung semantic segmentation and radiology report generation.The semantic segmentation of the lungs and the generation of radiology reports can help doctors accurately locate lesions,scientifically diagnose patients’conditions,and effectively alleviate problems such as doctors’ work pressure and imbalance of medical resources.However,current research faces the challenges of low model inference performance and low data utilization in lung semantic segmentation and radiology report generation.Specifically,the low performance of algorithm inference means that the semantic segmentation model has a large number of parameters and a large amount of computation,which leads to the slow inference speed of the model on hardware with limited computing resources;the low data utilization rate means that the radiology report generation model relies too much on image-text paired data.At the same time,there is also the problem of too long training time,which hinders the practical application of related technologies in CXR image analysis and processing.This paper optimizes related technologies to address the problems of low inference performance and low data utilization of the above-mentioned algorithms,and the main work is as follows:First,this paper designs a lightweight lung semantic segmentation model SegEnet based on Transformer.Consider both the segmentation effect of the model and the performance of model inference.The parameters of the designed model are only about 0.5M,and the amount of computation is about 2.7GFLOPs.Compared with the larger model,it has a good lung semantic segmentation effect and realizes efficient model inference performance in the model architecture.In addition,for the existing SegEnet model,to further improve the model inference performance in engineering,a variety of inference optimization strategies are used to improve the inference speed of the model.In particular,based on the open-source neural network inference engine,this paper has completed the hardware optimization for the MacBook using Apple Silicon,achieving a 2.05 times increase in inference speed.Secondly,this paper designs a radiology report generation model based on a memory enhanced Transformer.Using the visual Transformer pre-trained by self-supervised technology as the visual encoder of the model,the entire model can use a large amount of existing public CXR image data,reducing the algorithm’s excessive dependence on image-text paired data,and achieving efficient use of data.In addition,introducing the memory network module improves the quality of radiology report generation effectively.Finally,based on the above two algorithms,this paper constructs a Transformerbased CXR image analysis application system,which provides two functions of lung semantic segmentation and radiology report generation.It provides a web front-end interface for users to use.