Research On Self-supervised Liver Ultrasound Image Target Tracking Algorithm Based On Siamese Network

The tracking of liver ultrasound images mainly refers to the technical means of automatic tracking of liver lesion object points in ultrasound sequences through digital image processing technology.It is an important task in medical image analysis and is widely used in the evaluation of liver function,diagnosis of liver disease,and guidance of liver radiotherapy.However,due to the low resolution,noise,and deformation characteristics of ultrasound images,as well as the complex structure and motion patterns of the liver itself,relying on traditional manual feature tracking methods often cannot meet practical needs.Although supervised deep learning tracking algorithms can effectively improve tracking performance,they rely on a large amount of annotated data,which is difficult to obtain for liver ultrasound images.In addition,how to stably identify and locate objects for a long time in complex ultrasound image scenes is also a major challenge in current ultrasound image tracking tasks.This thesis proposes a self-supervised cycle-consistent object tracking method,SiamDCF,for liver ultrasound images based on the problems of insufficient data annotation and data distribution difference.The method learns the feature representation of liver ultrasound images by introducing a cycle-consistency loss function and performing forward tracking and backward verification of objects between sequential frames,and applies this feature representation in a Siamese network to construct a self-supervised tracking framework.The proposed method is validated on the CLUST(Challenge on Liver Ultrasound Tracking)2D dataset,and the experimental results show that the tracking error of the proposed method is2.01±3.11 mm,therefore demonstrating its effectiveness.This thesis introduces a context-aware correlation filter network,Siam-CCF,based on the self-supervised tracking framework to address the problems of noise,artifacts and edge area blur in ultrasound images,as well as the challenge of long-term tracking.The network uses the background information around the object as context,and dynamically adjusts its suppression weights according to the degree of interference caused by the context information to the object,thus effectively reducing the interference of non-object areas and improving the robustness of the algorithm in liver ultrasound tracking scenarios with background interference and noise.Furthermore,to prevent object loss or template contamination during long-term tracking,this thesis proposes a feature fusion strategy based on the object template.This strategy can adaptively select and fuse template features,which better accommodates changes in object appearance.The proposed method is validated on the CLUST 2D dataset,and the experimental results show that the tracking error of this method can be reduced to 0.79±0.83 mm.The liver ultrasound image object tracking algorithm proposed in this thesis is experimentally validated on the CLUST 2D dataset,and it achieves the fifth-best performance on this dataset,while achieving a good balance between accuracy and speed,making it one of the best methods.The work of this thesis provides a new idea and method for the field of liver ultrasound image object tracking,and therefore hopes to stimulate new discussions and thoughts on related research.