CT Thoracoabdominal Organ Detection Based On Classification Score Rearrangement And Convolutional Neural Network

In clinical medicine,it is very common to use computed tomography(CT)to assist diagnosis and treatment.It can help doctors to find out the disease area and problems quickly and accurately by obtaining the body information of patients by computer fault image technology,arrange treatment plans as soon as possible,and help them get rid of the trouble of disease.With the development of computer technology and artificial intelligence technology,many attention has been paid to the medical image processing by computer to help doctors to quickly and accurately detect the cause of patients.In order to ensure the correctness of subsequent diagnosis,the focus of computer attention should be focused on the diseased organs,which requires rapid and accurate location of organs,In this way,the influence of other health organs on the diagnosis results can be reduced and the accuracy and efficiency of doctors’ diagnosis can be improved.Therefore,a new method of detection of chest and abdomen organs based on classification fraction rearrangement and convolution neural network is proposed.Firstly,the current organ detection algorithm is improved.Because the current organ detection algorithm is usually a single stage organ detection algorithm,it is to directly obtain the organ category and boundary frame in the image after the feature extraction of medical image.This will lead to unbalanced positive and negative samples,which may lead to the detector not detecting some organs,In this paper,the single-stage detection algorithm is changed to a two-stage detection algorithm.The first stage produces the possible areas of organs.In the second stage,the regions are classified to find out which organs belong to,so as to ensure that the organs can be detected.Secondly,in order to reduce the mismatch between different size,different length to width ratio organs and preset anchor points,and reduce the over parameters of network design,this paper improves the regional recommendation network based on anchor point,and proposes a regional recommendation method based on thermal diagram.The advantage of the improved method is that the network does not need to be different in size,The size and aspect ratio of anchor points are preset for different aspect ratio organs,but the probability that the current characteristic points are organ center and the distance from the current point to organ boundary frame are directly returned,which reduces the over parameters of network design.Thirdly,because the classification accuracy determined by the neural network characteristics is higher than the regression precision,there is a great difference between the detected organ boundary frame and the real boundary frame.In this paper,the recommended area can be rearranged by classification score,so that the better region of the boundary box regression can be selected in the second stage,and the detection accuracy of the detector will be increased.Fourth,the regression speed of organs with different volume caused by the use of regional recommendation network based on thermal diagram is different,which leads to the increase of network training time.This paper adjusts the weight of loss function,which makes the larger organ gain greater weight,helps the network converge quickly and finally obtains the accuracy improvement in the data set.