Research On Lung Nodule Detection Based On Deep Learning

According to the research,pulmonary cancer is the leading cause of cancer-related death all over the world.In 2015,over 4.3 million were diagnosed with cancer in China,and lung cancer accounted for about 20% of it.Early detection of pulmonary cancer is the most promising way to improve patients' survival rate,patients diagnosed with pulmonary cancer have a five-year survival rate about 15%.However,if lung cancer is diagnosed in early stages,the five-year survival rate increases to 70%.Early lung cancer often appears as pulmonary nodules on CT images,so the automated detection of lung nodule in computed tomography(CT)images plays an important role in diagnosing pulmonary cancer.In some developed countries,such as America and Japan,high-risk citizens are suggested to take routine medical examination with low dose helical CT scanning.But it will produce hundreds of thousands CT images,and greatly increased the burden of radiologists.So there is an urgent need to use computer to help the radiologists to make diagnoses,and academic research in this area has always been a hot spot.Most of the previous methods of computer-aided lung nodule diagnosis are based on traditional machine learning.They extract feature manually,are hard to generalize and the sensitivity of them can still be improved greatly.So there is still a gap between theory and clinical application.Inspired by the success of deep learning in image processing,NLP,etc.,a method based on U-Net for lung nodule detection is proposed in this paper.U-Net is a convolutional neural network with an encoder-decoder structure,and it is commonly used in medical image segmentation,which input 2D image and make pixel level classification of them.This paper proposes some improvement on the U-Net.The structure of U-Net is adapted to input 3D images,and residual blocks are added into the neural network to improve its performance.The method of lung nodule detection includes three steps.The first step is image preprocessing.In this step,all raw CT data is converted to Hounsfield Units(HU).The second step is lung segmentation.In this step,threshold segmentation method and some morphological operations are used to separate lungs from other tissue,and extract lung maskimage.The last step is training a 3D U-Net to detect lung nodule.The CT images from step2 are sliced into 64 × 64 × 64 patches,and imputed into the network as 3D tensors to extractfeatures,learn from them to detect lung nodule and output the final result.LIDC/IDRI dataset,which is publicly available on internet,are used in this paper's experiment for training and testing.The dataset is split into 2 parts,90% of it is used for training and 10% for testing.It achieved 92.9% sensitivity with only 4.87 false positives per scan.The experimental result shows that the method of lung nodule detection proposed in this paper is better than the methods based on traditional machine learning,and can achieve the desired result.