Automated Segmentation And Identification Of Pulmonary Nodule Images

Lung cancer is one of the most common and fatal malignant tumors.The International Association for Lung Cancer Research(IASLC)International Staging Project confirms that the survival rate decreases with the increase of tumor size,indicating that early diagnosis and timely treatment are effective and key methods to reduce the mortality rate of lung cancer patients.CT plays an important role in the detection and characterization of pulmonary nodules.It is important to differentiate benign nodules from malignant nodules for guiding further clinical treatment.In recent years,with the development and application of CT,it can improve the sensitivity,specificity and accuracy of early diagnosis of lung cancer.Because medical images have the characteristics of blurred edges,uneven gray levels,large individual differences,serious noise and artifacts,it is difficult to achieve high sensitivity and accuracy in the research of related algorithms.According to the anatomical characteristics of thoracic CT images,a lot of experiments have been carried out in the two directions of pulmonary parenchyma segmentation,suspected pulmonary nodule detection and segmentation.The main work of this paper includes:(1)Automatic segmentation of lung parenchyma imagesIn order to improve the sensitivity,specificity and accuracy of the diagnosis of early pulmonary nodules in lung cancer,the lung image should be pre-processed,that is,the lung parenchyma segmentation.In this paper,an Automatic Anatomy Recognition(AAR)methodology based on fuzzy modeling idea and an Iterative Relative Fuzzy Connection(IRFC)delineation algorithm is studied,and lung parenchymal segmentation is performed on lung CT images.Fuzzy set concepts have been used extensively otherwise in image processing and 3D visualization.Fuzzy modeling approaches allow bringing anatomic information in an all-digital form into graph theoretic frameworks designed for object recognition and delineation.The uncertain information of the image is captured and encoded into the model.The methodology consists of five main steps: gathering image date for both building models and testing the AAR algorithms;formulating precise definitions of each organ in the thorax and delineating lungs following these definitions;building hierarchical fuzzy anatomy models;recognizing and locating lungs with the hierarchical models;delineating the lungs following the hierarchy.In recent years,clinical radiology research and practice have become more and more quantitative.In addition,the size and volume of the image are increasing.In order to make quantitative radiology feasible,it is essential to implement image segmentation algorithm and its fast implementation,and to generate real running time on very large data sets.In order to meet the practical running time,the AAR algorithm is optimized to improve the running speed of the algorithm.(2)Detection and segmentation of suspected pulmonary nodulesAfter pulmonary parenchyma segmentation,pulmonary nodules need to be automatically detected and segmented.This work is the premise of accurately extracting pulmonary nodule features,especially for the reduction of false positives.The main basis of automatic detection and segmentation of pulmonary nodules is that the CT value of the central area of the nodules is high and the boundary is irregular.In this paper,an improved Mask-RCNN based method for suspected nodule segmentation is proposed.The ResNet-based and DenseNet-based feature pyramid networks(FPNs)are modified according to the pulmonary nodule images,respectively,and are used as the backbone networks of Mask-RCNN.Based on the pixel-level prior knowledge of nodules,a pyramid feature map suitable for nodule images is extracted by Mask-RCNN,and then the region of interest(ROI)is obtained by the nodule boundary box provided by Mask-RCNN for nodule recognition and segmentation,which realizes the adaptive high-quality segmentation of CT images of pulmonary nodules.