A Deep Learning-Based Model For Detecting Lesion Signs In Lung Diseases

CT Signs mean the information of comprehensive manifestations of diseases at different pathological stages and levels.Automatic analysis of CT images outputs the locations and size of lesion regions which can help radiologists to make a credible diagnosis and effectively improve the speed and accuracy of clinical diagnosis.Sign-oriented research for lung disease detection can automatically analyze the location and size of the output lesion area to help radiologists make a credible diagnosis,effectively improving the speed and accuracy of clinical diagnosis.The paper fills the research gap of target detection tasks with signs as the starting point,which has important academic significance and clinical application value.The main research and innovation points of the paper are as follows.This article provides a systematic review of the current research status of lung CT sign detection from the perspective of sign detection by integrating current research from both domestic and foreign researchers to fill the research gap in this field and provide a reference for scholars.The article systematically analyzes and compares the traditional machine learning methods and deep learning methods in the detection of lung CT signs,concludes the current problems and difficulties to be overcome in this field,determines the improvement direction of future research work,and provides a basis for later research.To address the problems of low accuracy and poor robustness of current target detection methods for detection of small lung lesion signs,a new end-to-end improved small target detection network based on YOLO is proposed to improve the accuracy of small target detection by means of small target expansion and the designed path-enhanced aggregation network structure,and the effectiveness of the proposed method under small target detection is experimentally verified,which provides a solid research basis for later migration to lung CT images for detection of focal lesion signs in lung lesions.Experiments show that the best m AP(mean Average Precision)of our method is 99.45%for small target detection in natural images and 66.06% for lung CT sign detectionTo address the current problem of lack of well-labeled sign datasets for training deep learning networks,a new semi-supervised learning-based lung CT sign detection model is proposed based on the proposed end-to-end lung lesion sign detection network,and the designed teacher-student mutual learning mechanism can correct pseudolabel in a timely manner,effectively ensure the quality of pseudolabel,and significantly improve the accuracy of the model,and improves the m AP of CISLs detection from 66.06% to 94.36%,outperforming the supervised methods by 28.30% on the publicly available LISS dataset.