Reserach On Semantic Feature Of Pulmonary Nodule In CT Images Based On Causal Structure Learning Of Streaming Features

Lung cancer is one of the fastest growing malignancies with increased morbidity and mortality.Early diagnosis is essential for lung cancer treatment.At present,clinical tomography mainly relies on computer tomography to screen for pulmonary nodular lesions.Computer-aided methods based on medical images to detect and diagnose are of great significance for the treatment of cancer patients.Due to the complex lesion area and various representations in CT images of lungs,physician screening has the disadvantages of heavy workload and obvious differences between observers.How to effectively extract the features of the lesions and use the relevant features to classify the semantic attributes of lung nodule images is currently One of the research hotspots in the field of biomedicine and machine learning.This paper mainly studies the classification of semantic attributes based on flow features in CT images of lung nodules.The specific work is as follows:Firstly,feature extraction of lung nodule lesions is a very important part of the experiment.Due to the complex and diverse clinical representation of lung nodule lesion areas in CT images,the calculated feature set should describe the lung nodule lesion area more comprehensively,and the problem of feature redundancy between the calculated features needs to be avoided.In this thesis,the underlying two-dimensional image features will be used to represent the lung nodule lesion area.Secondly,for the imbalance problem of the feature dataset,we use the smote algorithm to solve the data imbalance problem.In addition,there is no uniform dimension page in the data set that will affect the classification results of the experiment.Based on this,we discretize and normalize the feature set.Then,due to the common semantic features of CT images of lung nodules,it can provide rich quantitative clues for the in-depth analysis of lung nodules.Aiming at the mapping relationship between the underlying computing features and semantic features of lung nodule images,we first propose a causal structure learning algorithm(CD-SF)based on the chi-square test for stream features.The CD-SF algorithm A feature is processed and screened in time to eliminate irrelevant features at the earliest possible stage,create a smaller subset space for subsequent conditional independence tests,and also create a smaller search space for the search orientation stage,and then establish Causal networks perform hierarchical classification of semantic attributes through joint tree prediction algorithms.Finally,in view of the lack of time performance of the CD-SF algorithm,we propose the CD-SU-SF algorithm.In this algorithm,for each feature generated one by one,we calculate the mutual information and set the correlation threshold The possible candidate neighbor nodes were obtained,and the exponential computational complexity was optimized to a linear level.Experiments show that the algorithm also achieves good timeliness and accuracy.