Detection And Application Of Hepatic Portal Hypertension Based On CT Images

One of the significant predictive factors for cirrhosis is clinically significant portal hypertension.However,the gold standard for diagnosing portal hypertension is portal venous puncture,which is invasive and only available in specialized centers,hindering its routine use in clinical practice.Therefore,a non-invasive,accurate,and reliable method for diagnosing portal hypertension is urgently needed in clinical practice.This study focuses on the detection of portal hypertension based on CT images.The difficulty of detecting portal hypertension based on CT images lies in the fact that abdominal CT images contain many organs,and redundant information can interfere with the model’s fitting of key features.To address this issue,this study proposes a research approach of first segmenting and extracting the key target organs and then studying portal hypertension detection based on target organ features.The main research work of this paper is as follows:(1)Organ segmentation based on fully supervised learning.Portal hypertension usually causes splenomegaly and portal vein dilation.To eliminate interference from other organs,this study targets the spleen and portal vein for segmentation.To address the low contrast and large differences in the segmentation target region of CT images,a multi-scale attention mechanism is designed to make the model pay more attention to the spleen and portal vein regions.To address the poor edge segmentation performance of the original network,a fusion downsampling module is designed to make the model more robust to boundary information.(2)Organ segmentation based on semi-supervised learning.Supervised learningbased organ segmentation requires a large amount of annotated data,and medical image annotation often requires professional knowledge,greatly increasing the annotation cost and making fully supervised methods difficult to promote.Semi-supervised learning can reduce dependence on annotated data.The semi-supervised segmentation network designed in this study uses two CNNs with different sizes of convolution kernels for cross supervision.To fully utilize the encoding samples generated by the dual CNN architecture,a contrastive learning module is designed.(3)Portal hypertension detection with fused target organ information.Abdominal CT images contain many organs,and to reduce noise interference from other organs,this study obtains the segmented liver and spleen through the designed segmentation network and sets these two organs as target organs.Feature fusion is then performed with CT images to enhance the classification features in a targeted manner.Through a large number of experiments,it has been shown that CT images with fused target organ features can better accomplish the task of portal hypertension classification.