Automatic Lung Parenchyma Segmentation And Pulmonary Nodules Detection Based On Deep Learning

Lung cancer creates an estimated 1.3 million deaths annually and that is the primary contributor to cancer deaths all over the world.As an early form of lung cancer,pulmonary nodules can be diagnosed and treatment initiated by experienced physicians observing their location,extent and shape in chest computed tomography scans.However,many unstable factors,such as distraction,fatigue,and limitations of professional experience,may lead to misjudgment by the physician and affect the subsequent treatment plan.Therefore,using the computer aided diagnosis（CAD）for the therapy and longer the survival of patients is of great significance.As two key steps of computer aided diagnosis,lung segmentation as well as pulmonary nodules detection can be performed using deep learning approaches instead of the previous conventional machine learning algorithms.Unlike the previous algorithms,current in-depth learning model are not only automated highly without predefined,but also widely used in various CAD systems.Recently,medical image processing based on deep learning provides a critical research direction for medical image analysis.Thus,in this paper,deep learning is adopted in the tasks of lung parenchyma automatic segmentation and pulmonary nodule detection.The main work is as follows.（1）To further improve the segmentation results,a convolutional neural networkbased lung segmentation method is proposed in this paper.Firstly,we propose an advanced method with encoder-decoder.In order to avoid background interference and enhance the model ability to learn,the attention mechanism is applied to the decoder.Secondly,to gain multi-scale contextual information sufficiently,we deploy the DAC block between the encoder and the decoder,which inherits the advantages of Inception,Res Net and atrous convolution.Finally,up/down sampling modules with different convolutional kernels are utilized to widen the network.Based on a series of ablation and comparative experiment,we found that the Dice similarity coefficient of our algorithm is0.9859,whose performance is better than other prevalent segmentation algorithms.Therefore,it is an important preprocessing step to separate lung parenchyma from the CT images automatically and precisely.（2）Since lung nodules are small in size and vary greatly in size,they belong to small target detection,therefore,a Two-stage algorithm for multi-scale detection is designed in this paper.First,the algorithm will use Res Net network as the backbone network for feature extraction.Besides,in order to reduce the difficulty of feature extraction,the highlevel features are combined with the low-level features one by one while feature extraction to form a feature pyramid network to enhance the semantic features of each layer of the network.Finally,the cascaded R-CNN network is used to improve the overall detection accuracy by setting different thresholds to enable the network to adapt to different scales of targets.This paper performs the experiment on LUNA16（Lung Nodule Analysis 2016）dataset.Meanwhile,the distribution of small nodules in the training set is enhanced by Mosaic data augmentation,which makes the model robust in small nodule detection.The improved algorithm proposed in this paper for multi-scale lung nodules can effectively solve the problem of difficult detection of lung nodules of different sizes,and its detection accuracy is improved for targets with smaller areas significantly.