Key Techniques For Lung Nodule Detection And Classification Based On Chest Imaging

Lung cancer is one of the most common cancer-related diseases in the world,and is considered the leading cancer in causing death worldwide.Early detection and diagnosis of lung cancer can greatly improve patient treatment during the early stage and hence potentially increase the survival rate and patients’ quality of life.Early stage lung cancer manifests itself as pulmonary nodules.Therefore,the early stage detection of lung nodule can potentially reduce lung cancer mortality.X-ray is one of the most commonly used imaging modality for chest radiography.Due to its low cost,low-dose,and penetrating ability,X-ray has been widely used in clinics to diagnose pulmonary diseases.Because different anatomical structures absorb varying levels of X-ray,the projected image of the anatomical object of interest is obtained.However,it is very difficult to detect the lung nodules because their variable size and grayness,as well as overlapping by the human skeletal rib.Therefore,an accurate detection of lung nodules is crucial to increase detection of the early stage lung cancer,which has recently become an active research area.Computed tomography(CT)is very popular imaging procedure in the daily radiological practice.Now,thin-slice helical chest CT scans have a sub-millimeter resolution at which small pulmonary nodules can be detected.However,due to the large number of slices,the rich structure of vessels and airways in the lungs complicates the search task significantly.Since the shortcomings of traditional manual chest imaging in the diagnosis of chest disease such as heavy workload,long reading priod,and strong subjectivity,lung nodule computer-aided diagnosis systems are needed with the function of rapid,accurate and repeatable to assistant doctors in the diagnosis of diseases.Based on the study of recent and relevant literatures,we have performed a complete research on lung nodule detection and classification in chest imaging(both X-ray and CT images).The main contributions of this dissertation are as follows:(1)Proposing a deep feature fusion method to classify the lung nodule in chest radiographyWe first attain identification of the suspicious lung nodules using the generalized Laplacian of Gaussian Filter.Then,we extract the deep feature fusion from the nonmedical training and hand-crafted features.Our results show that,the deep fusion feature can achieve good results in terms of sensitivity 0.69 on the public dataset JSRT.(2)Proposing an automated chest screening based on a hybrid model of transfer learning and convolutional sparse denoising autoencoderIn order to classify the level of the chest radiography,we combine the probability confidence from convolutional neural networks(CNN)classifiers and the reconstruction error from the convolutional sparse denoising autoencoder to obtain score function.Finally,we vote for the decision of the level of the image.The experiments show that we can achieve the high accuracy for classifying these images up to three levels.It will allow radiologists to focus their attention immediately on higher-risk cases and correct potential misdiagnoses.Our method yields an accuracy of 0.7,recall of 0.74,and F1 score of 0.72 on the abnormal data,while it achieves an accuracy of 0.9,recall of 0.9,and F1 score of 0.91 on the normal case.Compared with existing experiment results,our method achieves promising results in terms of accuracy,recall and F1.Compared with the ground truth,we can achieve precision of 0.99(390/395)on the normal case,0.94(83/88)on abnormal case.(3)Proposing a deep regression for landmark detectionsIt is of great importance to accurately locate the skeletal ribs for subsequent rib suppression process.We propose a deep regression for landmark detection to obtain the robust sampled landmarks.We can accurately identify the location of the ribs from these regression points based on the active shape model.Our experimental results show that,on the test set we can achieve a good result with an average IoU of 0.76 and an average Dice coefficient of 0.83.(4)Proposing a 3D deep neural network for lung nodule detection and classification from CT imageWe detect the lung nodule region using a similar 3D U-net,and then obtain the features from the suspicious lung nodules from a feature pyramid network.The sensitivity of the algorithm proposed in this paper reaches 0.84.Based on the 3D spatial context information,almost best results are achieved in term of sensitivity when compared with other state-of-the-art methods.