Automatic Calibration And Segmentation Of Medical Image Target Area Based On U-Net

Medical images in quantitative research and plays an important role in medical diagnosis,but traditional radiology department for CT,MRI,X-ray medical image such as the analysis is mainly rely on artificial try to interact,in the treatment of many diseases in the process of doctors need to be based on patients medical image outline the lesion area were analyzed,and the efficiency of the traditional way in a clinical environment has always been relatively low,at the peak of hospitalized very prone to delay treatment of best time.At present,many achievements have been made in the segmentation of CT and X-ray.However,based on the current mainstream segmentation methods,the redundant use of information is resulted,and similar low-level features are extracted many times on multiple scales.Secondly,the long-term feature dependence is not effectively modeled,which leads to the non-optimal discrimination feature representation related to each semantic class,and the semantic segmentation caused by the excessively large network has a certain gap with the industrial application in effect or speed.The main purpose of this research is to try to overcome these limitations with the proposed architecture,to reduce redundancy and improve performance by capturing richer context dependencies based on the auto attention mechanism,and to speed up the compression of the model so that it runs on edge devices.Aiming at the problem of lung pneumothorax data segmentation,this paper explores a method that can integrate local characteristics and their corresponding global dependencies and highlight the interdependent channel mapping in an adaptive way by referring to the common full convolutional neural network such as u-net,which has become the field of medical image segmentation.Finally,the deployment of the model on the edge equipment is completed,and the model is quantized from Float32 to Uint8 by means of quantization perception training for the model,and the DSC of the model is guaranteed to have no obvious drop,so that the segmentation task can be completed conveniently,efficiently and accurately.By emphasizing the correlation of features to enhance the additional loss between different modules,the attention mechanism is guided to ignore irrelevant information and focus on more discriminating areas in the image.Secondly,in order to solve the problem that the segmentation model is too large and the industrial deployment is difficult,based on the strategy of channel mixing on the basis of deep reelable volume,the information communication and computation compression are strengthened,and the short connection structure is added to the main network of the network to make the network easier tofit.Then,pruning is carried out on the model to compress the number of model parameters,reduce the number of visits and the amount of computation,so as to lower the deployment threshold of the model and accelerate the inference.Finally,the deployment of the model on the edge equipment is completed,and the model is quantified from Float32 to Uint8 by means of quantization perception training for the model,and the DSC of the model is guaranteed to have no obvious drop,so that the segmentation task can be completed conveniently,efficiently and accurately.