Research On Classification Algorithm Of Standard Plane Of Ultrasound Liver Based On Convolutional Neural Network

Medical imaging(Medical Imaging)technology has been widely used in clinical examination,diagnosis,treatment and decision-making,and has gradually become an indispensable important tool for clinical diagnosis and treatment and surgical planning.Among them,the ultrasound-based examination method is widely used in clinical diagnosis and treatment because of the advantages of convenient examination,low examination cost,and non-invasive trauma and safety to the patient during the examination process.Clinically,the ultrasound examination process is generally divided into the following three steps:(1)the clinician selects the corresponding inspection angle according to the organ to be examined,and takes a standard slice image of the corresponding angle;(2)measures the features in the standard slice image,The measured features include the length of the organ,the thickness of the blood vessels,etc.(3)The clinician makes the final clinical diagnosis for the patient based on the measured features.Among them,the standard section images found in the clinic are a set of images that can clearly describe the structure of the examined organ.This set of standard section images is helpful for doctors to understand the health of the entire organ from various angles in detail and help to make the final Diagnosis.However,the organs of different patients have different characteristics,such as the size of the organs and the distribution of blood vessels.Therefore,the positioning of standard section images requires the clinician to have a wealth of clinical experience and a complete knowledge reserve.At the same time,this is also a very time-consuming task that requires careful search by the doctor.In view of the above problems,if computer aid can be introduced in the clinic,the diagnostic efficiency of ultrasound will be greatly improved.Therefore,this paper designs a related auxiliary diagnosis algorithm based on convolutional neural network.The main research contents are as follows:Ultrasound can check most organs of the human body.But among them,the ultrasound of the liver is the most extensive.Therefore,this paper chooses to design eight classification algorithms for eight standard sections of the liver.The main function of this algorithm is to classify the input ultrasound image and determine which type of standard section it belongs to,thereby improving the efficiency of doctors in locating the standard section.The algorithms in this paper are mainly divided into ROI segmentation algorithms and classification algorithms.The specificdetails are as follows:This paper implements a Dense-U-Net-based algorithm for segmenting regions of interest in ultrasound images.This segmentation algorithm mainly extracts the intra-class specific features of standard sections of different classes.Dense-U-Net is an improvement on U-Net.It changes the connection mode between each convolutional layer to dense connection,so as to improve the feature reuse rate and improve the segmentation result.The entire algorithm process first cuts each picture into 48 48 small pieces,and then uses the cut pictures for training.Through experiments,Dense-U-Net can extract the ROI of pictures very well.At the same time,for the input pictures before and after ROI extraction,the classification result of the final classification network has a 5% improvement effect.This paper implements an ultrasonic image classification algorithm based on transfer learning.The collection of ultrasound images is difficult and time-consuming,so the amount of data is small.However,convolutional neural networks trained on small data sets are prone to overfitting problems.Therefore,this paper establishes a standard liver section classification based on transfer learning.The training process first uses the large public data set ImageNet to pre-train the classification model,and then uses the liver ultrasound data set to fine-tune the fully connected layer of the network.The network includes 16 convolutional layers and 3 fully connected layers.The experimental results show that the method of transfer learning can well overcome the problems caused by insufficient data volume.Finally,experiments prove that the network can reach a classification accuracy of 90.35% in the test set.This paper implements an ultrasound image classification algorithm based on the attention mechanism.In order to further improve the operation effect of the classification network.We consider that the intra-class features of the standard aspect have a certain discrimination,and different intra-class features have different scales.Therefore,we increase the attention mechanism in the network,and combine the image features of different levels and different scales to make the final category prediction for the image.The convolutional layers of the classification network are connected by residual connections.The entire network includes 13 convolutional layers and 3 fully-connected layers.The input of the fully-connected layer combines 3mesoscale features.The experiments show that the classification accuracy of the model in the test set reaches 94.36%,which is 4% higher than the previous network.