Lung Nodule Detection With Deep Learning

Early detection of lung cancer is helpful for curing the disease and reducing the mortality rate.The most important early symptom of lung cancer is the pulmonary nodule.So,the detection of pulmonary nodules is crucial for curing lung cancer.The low-dose Computed Tomography(CT),which can help doctors detect lung nodules,is important for the successful diagnosis and treatment of lung cancer.Currently,radiologists analyze CT scans to check pulmonary nodules.However,due to the large number of patients and the large size of CT scans,radiologists have to check so many CT scans every day that they undertake heavy burdens.Also,there are a lot of inconspicuous nodules that the radiologist may make mistakes.With the image processing technology,computer-aided systems(CAD)can automatically detect candidate nodules and provide them to radiologists that it greatly reduce radiologists' pressure and improve the quality and efficiency of nodule detection.Compared with traditional machine learning methods,deep learning has the ability to automatically extract features and shows superior performance in image recognition.So,we also implement the detection and classification of pulmonary nodules through deep learning.We trained and validated the deep learning model using the large publicly available dataset,Lung Image Database Consortium(LIDC-IDRI).We obtain the CT scans information from the LIDC-IDRI dataset,preprocess them by extracting the lung lobes with traditional image segmentation methods,and then conduct the resampling and normalizing processes on the data.Also,we enlarge the dataset by performing rotation,randomly cropping and scaling.This paper proposes the selection of candidate nodules based on the U-Net model.By adjusting and optimizing the U-Net network structure,the sensitivity of the candidate nodules is enhanced and the detection rate of nodules is increased.At the same time,many false positive nodes are also added to the pool of candidates.To distinguish true-positive nodules from false-positive nodules,a false-positive screening is required.In this paper,we propose the false positive screening based on three-dimensional convolutional networks.We also studied and analyzed the influence of different structures and parameters and achieved 93.2 percent sensitivity.In order to obtain a higher detection sensitivity of false positives,we create a new network by combining multiple branches of the residual network with multiple shortcut connections.Without increasing the number of parameters,our new three-dimensional residual network achieves 94.5 percent sensitivity.