Lung Segmentation Based On Convolutional Neural Networks

Lung segmentation is a prerequisite for examination of pulmonary function and diagnosis of pulmonary disease.In recent years,although researchers have proposed lots of methods on lung segmentation,these methods usually have its own limitations,which aimed at the specific thoracic CT images or specific disease of pulmonary disease using manually extracted features.Thus,with minimal human intervention,we need to find a way which is more robust to extract lungs fast and accurately.In this paper,first we introduce the development of deep learning,which includes the basic concepts and common neural networks,and implement analysis with the preprocess of image data.Based on previous theories and works,this paper proposes a model framework of lung segmentation using deep learning,including image preprocessing and lung segmentation based on convolutional neural networks.In details,image preprocessing consists of image grayscale transformation,image histogram processing and image denoising,which mainly to ensure the consistency of brightness and contrast when using different devices or parameter settings to obtain thoracic CT images.In the stage of lung segmentation,our task includes construction,training,parameter tuning and results evaluation of the lung segmentation neural networks.We use fully convolutional neural network to implement lung segmentation,which takes the preprocessed chest CT images and corresponding masks as inputs during training.In the period of training,we conduct fine tune to important parameters.Finally,we get a deep neural network model which extracts features from data automatically to segmented lungs by classifying the pixels of chest CT images.Eventually,mapping the classified results to binary image,which is the corresponding segmentation result of original chest CT image.We use the final model to segment lungs on test data,the average Dice coefficient of lung segmentation results achieves 0.9894,which is stable,fast and precise.