Real-time Tracking Method Of Soft Tissue Target In Lung

Real-time accurate positioning of lung lesions has always been a difficult problem in the diagnosis and treatment of lung diseases.Owing to the dynamical movement of lungs soft tissue with interference factors such as breathing and heartbeat,traditional CT static images or in vitro markers-based positioning methods can only provide information about the location of important tissues such as lesions,blood vessels,and trachea at a specific time,and it is difficult to obtain accurate information.Dynamic location information,which brings great difficulties to lung tumor puncture pathological diagnosis and surgical treatment.In order to solve this problem,this paper studies the lung soft tissue target tracking method under breathing motion.With the help of real-time observation data of lung fluoroscopic video,a method for tracking spatial movement features of lesions based on machine learning algorithms is proposed to solve the problem.Since lung tissue in fluoroscopic is usually difficult to identify,this paper uses the idea of tracking sparse lung markings features to estimate lung tissue movement information.For lung texture,it is difficult to train a deep network model due to the lack of large public data sets.At the same time,the lung texture image will be disturbed by noise and it is difficult to use traditional features to extract.Therefore,this paper proposes a feature matching method that uses trans-fer learning as a feature extractor and a continuous spatial domain correlation filter as a feature matcher to track lung texture.The method includes two main contents:constructing a convolu-tional neural network to extract texture features and dynamically matching features.1)Aiming at the problem of insufficient lung tissue image information,this paper proposes a method based on transfer learning as a deep and shallow activation feature extractor of neural network,which can effectively identify lung markings features.This paper uses the transfer learning method to get over the lack of large medical data sets.At the same time,the deep and shallow layers of the con-volutional neural network are used to solve the noise interference,and the lightweight network is used to speed up the feature extraction speed.2)Aiming at the problem of target tracking,this po-sition proposes a feature matching method based on the correlation filter in the continuous space domain,and improves its real-time performance and robustness by reducing the dimensionality and optimization of the filter model.This paper analyzes the characteristics of the tracking target,simplifies the model based on the basic correlation filter,reduces the frequency of learning,and achieves higher real-time performanceCompared with the existing lung target tracking methods,the method in this paper expands the types of tracking targets,and at the same time achieves a higher tracking accuracy.The al-gorithm in this paper has been experimentally tested on the fluoroscopy video of large dogs and the digital reconstructed image generated by the four-dimensional computed tomography.The av-erage tracking error of the lung tumor target is less than 1mm,and achieved with 14 frames per second on fluoroscopy video.The method in this paper can be realized with only low-cost fluo-roscopy equipment,and is suitable for many fields such as puncture biopsy,radiotherapy,surgical positioning and so on.