| Lung cancer is currently one of the malignant tumors with the highest morbidity and mortality in the world.Most patients have difficulty detecting early lung disease.When clinical symptoms appear,they are basically in the middle and late stages of the disease course,and the therapeutic effect is limited.Pulmonary nodules are an early manifestation of lung cancer.The detection of pulmonary nodules is of great significance for reducing the mortality of lung cancer and improving the survival rate and quality of life of patients.Computed Tomography(CT)images of the lungs are currently the main method for the diagnosis of early lung cancer.However,manual diagnosis has strong subjectivity and low consistency,which can easily lead to missed diagnosis and misdiagnosis.Deep learning can acquire deep-level features of images and has strong generalization ability.With its strong robustness,it can be effectively applied to medical image processing to realize intelligent analysis of lung CT images.It is an important trend to study computer aided diagnosis(CAD)system based on deep learning.The detection of pulmonary nodules in early lung cancer CAD system usually consists of three parts: lung CT medical image enhancement,candidate nodule detection and false positive reduction.Pulmonary nodules have variable shapes,different sizes and complex backgrounds in lung CT images,which make the detection of lung nodules have certain problems.During the lung CT imaging process,it is disturbed by factors such as environment,equipment,breathing movement,etc.,which results in blurring of details such as image edges and textures.In response to this problem,this paper proposes a CT image saliency area enhancement algorithm for pulmonary nodule detection.The algorithm realizes the truncation and normalization of the pixel values of lung CT images,and uses pixel-based gamma transformation enhancement to improve image quality.The maximum betweenclass variance method is used to separate the front and background,combined with morphological operations to complete the lung parenchymal repair,which is used to refine particles,smooth edges and fill blood vessels.Finally,it achieves refined segmentation of lung parenchyma and achieves significant area enhancement.The small size,random location,irregular shape,poor contrast and complex surrounding background of pulmonary nodules cause a large number of missed pulmonary nodules.In response to this problem,this paper proposes an end-to-end candidate nodule detection algorithm based on attention mechanism.The low sensitivity of lung nodules is due to the small size of pulmonary nodules and their low proportion in CT images.This paper studies the application of deep learning in small target detection,and analyzes the deep feature fusion strategy.The end-to-end bounding boxes prediction of candidate nodules is performed by using high resolution feature map,and the attention mechanism based on channel and space is introduced.Ultimately improve the sensitivity and accuracy of pulmonary nodules.Early CT imaging features of pulmonary nodules are unclear.It is difficult to distinguish the positive and negative pulmonary nodules based on CT images.In response to this problem,this paper proposes a false positive reduction algorithm based on multiscale feature fusion.The main reasons for the high false-positive rate of lung nodules are the large scale transformation of the nodules,the serious imbalance of positive and negative samples,and the difficulty in characterizing feature information.Therefore,this paper uses 3D convolution to extract rich context information of 3D nodules.Multi scale feature fusion is used to avoid the limitation of single scale classifier.Meanwhile,hard case mining and data enhancement algorithm are introduced to balance the positive and negative of data set.Ultimately improve the accuracy of nodule classification and reduce the false positive rate.Combined with the lung CT image enhancement,candidate nodule detection and false positive elimination studied in this paper,a lung nodule detection system based on deep learning is designed and implemented.It realizes the positioning,qualitative and quantitative of pulmonary nodules,and provides users with reference opinions for auxiliary diagnosis. |
