Detection Of Small Lesions In Liver CT Images Based On Improved Faster R-CNN

As the largest solid organ in the stomach of the human body,liver has many types of small lesions,and the incidence of liver disease is very high.This poses a serious threat to human health and life.Therefore,clinical diagnostic medicine research in China has always focused on the diagnosis,prevention and treatment of liver disease.In recent years,Computed Tomography(CT)has become the most widely used medical imaging method for the diagnosis and treatment of liver diseases.However,the size of the liver varies with gender,age,and body shape.And the boundary between small lesions and surrounding normal liver tissue is not clear.In addition,the diversity in size and appearance and heterogeneous density increase the difficulty of detecting small lesions in the liver.Therefore,the study of detection algorithm for CT images of small liver lesions is invaluable to guide clinical diagnosis and treatment.This paper presents an automatic detection algorithm of small lesions in liver CT images based on deep learning.The main research contents are as follows:(1)Since there are few existing datasets for detection of small liver lesions,small lesions on liver CT images are manually annotated under the guidance of specialists,and a highly accurate dataset of small lesions on liver CT images were constructed for the training and evaluation of the model.(2)To address the problem that the current mainstream detection models are poor at detecting small lesions,an improved Faster R-CNN-based algorithm for small lesion detection in liver CT images is proposed in this thesis.A network structure based on Res Net101 and Feature Pyramid Networks(FPN)is used to realize feature extraction from liver CT images,and high-level semantic information is integrated into the feature information of the underlying network;labels are assigned based on NWD(Normalized Gaussian Wasserstein Distance)metrics in RPN(Region Proposal Network).In addition,Ro I Align is used in the feature mapping layer to effectively reduce the deviation generated by feature mapping.Finally,comparative experiments based on Faster R-CNN are conducted using different enhancement strategies.The results of the experiment show that the improved Faster R-CNN algorithm greatly improves the Average Precision of detection of small lesions on CT images of the liver.(3)In order to improve the detection speed of small lesion detection model in liver CT image,this thesis improves the RPN network on the basis of the previous work by using the double-threshold segmentation method and Marr-Hildreth edge detection algorithm to extract liver CT image of small lesions corresponding contour boxes as pre-selected boxes,and the score of the preselected boxes are calculated,and the center point of the pre-selected box with a high score is selected as the anchor,so that the RPN network can generate targeted anchor boxes,which effectively alleviates the problem of unbalanced positive and negative samples,and greatly reduces the time and computational cost of model training.Finally,comparison experiments with the mainstream objective detection models are performed in this thesis,demonstrating that the detection algorithm based on the improved RPN not only outperforms the mainstream objective detection model in accuracy,but also has a great improvement in detection speed.Therefore,this thesis proposes an automatic detection algorithm of small lesions on CT images of liver based on deep learning.It has special application value for the diagnosis and treatment of liver diseases.