Hepatic Portal Vein Detection Based On Semi-supervised Learning

As an important indicator of human body,hepatic portal hypertension can easily cause death of patients,so it is necessary to detect hepatic portal hypertension in patients.At present,the traditional detection method of hepatic portal hypertension is surgical measurement,which will cause trauma to patients and is not efficient.In order to improve patients’ medical experience and detection efficiency,this paper carries out research on hepatic portal hypertension detection based on deep learning.The difficulties of hepatic portal hypertension detection based on deep learning are as follows: screening of medical raw data is required to meet the requirements of model training,and manual screening will be time-consuming and increase the economic cost;How to comprehensively analyze CT images and how to improve feature quality is the key to determine the performance of the model in the detection of hepatic portal vein hypertension.In view of the difficult problems,the main research work of this paper is as follows:(1)The data cleaning work based on semi-supervised learning is studied,in this study thoracic CT scan content is complex and there are many interference factors,which will make network feature extraction difficult.Therefore based on the structure of the hepatic portal vein position characteristics,the targeted attention mechanism is designed and added.It makes the model more focus on hepatic portal vein area,reduce interference with the rest of the area;In addition,for the normal sample imbalance problem,Focal Loss is also used to replace the original Cross Entropy Loss,so that the model can assign greater training weight to the categories with poor prediction.(2)Vision Transformer(Vi T)network with global receptive field is applied to the diagnosis of hepatic portal hypertension in view of the numerous characterization of hepatic portal hypertension.A detection method of hepatic portal hypertension based on Vi T is established.Since the original Vi T did not use convolution operation to extract texture information at the bottom of the image,therefore,the performance of similar target classification needs to be improved.Aiming at this problem,Contrastive Loss is added to the Loss function in addition to the original cross entropy Loss to enhance the network’s recognition ability of similar targets.Experiments show that the recognition result is significantly improved after the addition of the Contrastive Loss function.(3)Medical images usually contain very complex information,and it is difficult to fully express the content of medical images only by a single feature.In this paper,the pressure area feature is designed to fuse the depth feature with the pressure area feature,so that the fused feature has richer semantic information,excellent stability and interpretability.After a lot of experiments,it is found that compared with single feature,feature fusion can improve the recognition accuracy and convergence speed of the model.Finally,the detection system of hepatic portal vein hypertension is constructed by using this method.