Automatic Detection Of Lung Nodules Based On CT Images

Lung cancer is the most deadly cancer in the world today,with a 5-year survival rate of only 17.7%,but studies have shown that if an effective diagnosis is made early,the 5-year survival rate can be increased to 54.4%.Therefore,early detection,diagnosis and treatment of lung nodules are important tools to improve the survival rate of lung cancer patients.Early symptoms of lung cancer appear in the form of lung nodules,therefore,early detection and timely treatment of lung nodules are important to save the lives of lung cancer patients.CT imaging has become the main tool for physicians to diagnose lung nodules.Compared with previous imaging techniques,CT examinations obtain a large amount of image data,which can provide more information about organs and tissues,and also impose a great reading workload on physicians.In order to improve physicians’ work efficiency,reduce work intensity,overcome the influence of human factors in films,and improve the detection rate of pulmonary nodules,there is an urgent need to study efficient and accurate automatic detection methods for pulmonary nodules in lung CT images.The main work of this thesis is as follows.1)In lung parenchyma segmentation,this thesis combines morphology and threshold processing to first extract the thoracic mask and lung parenchyma mask,and for the problem of pleural subsidence caused by pleural adherent nodules,the closed operation in morphological filtering is used to repair;in order to remove the trachea of the lung,the left and right lung parenchyma is obtained by using the connected region analysis to take the first two connected regions with the largest area.The Dice similarity coefficient of segmentation measured on the co-common dataset was 0.9782,and the average segmentation time per CT slice layer was 0.6397 seconds.2)In terms of suspected lung nodule extraction,this thesis proposes a method based on mean drift clustering to detect suspected nodules,which can significantly increase the contrast between lung nodules and surrounding tissues,and subsequently extract suspected nodule regions using adaptive dynamic thresholding and morphological closure operations.3)In terms of reducing the false positive rate,this thesis extracts the grayscale,shape,and texture features of the study object from various aspects based on the imaging characteristics of lung nodules,and selects and fuses the multidimensional features to reduce the dimensionality of the features,uses the classifier to reduce the false positive rate of nodules,and verifies the importance of fused features to improve the accuracy of detection.In this thesis,we study lung nodule detection based on CT images,deeply study lung parenchyma segmentation,suspected nodule region detection and false positive rate reduction,and achieve automatic detection of lung nodules.We innovatively propose a lung parenchyma segmentation method combining threshold and morphology,which can segment lung parenchyma regions simply and quickly,and also propose a lung nodule candidate region extraction method based on mean drift clustering,which is useful for the subsequent In addition,a method based on mean drift clustering is proposed for extracting candidate regions of lung nodules,which can greatly improve the detection performance.