Research On Lung Image Segmentation Based On Deep Learning

X-ray chest X-ray is an important means of detecting lung diseases.With the increasing incidence of lung diseases,computer-aided diagnosis technology is of great significance for clinical treatment.However,X-ray chest images generally have image quality problems with unclear lung field boundaries and severe inter-tissue interference,which makes it difficult to segment lung tissue.The use of computer-aided diagnosis to identify chest X-ray images has become a hot research direction and an important trend,which can alleviate the uneven status of regional medical standards.With deep learning in the field of image processing,the semantic segmentation technology has achieved excellent segmentation performance in complex scenes from the continuous traditional convolutional FCN to the hollow convolution.Using the deep learning method to segment the lung tissue can not only solve the local convergence problem encountered by the traditional image segmentation method,but also improve the segmentation accuracy.This paper relies on the research project of pneumoconiosis classification and detection,and divides the lung tissue as one of the work of the project.The main research contents of this paper are to divide the lung tissue by improving the FCN and Deeplabv3+on the basis of summarizing the traditional image segmentation method to segment the lung tissue.The specific content is:(1)Summarize the deficiency of traditional image segmentation algorithm to segment lung tissue.The most representative ones in traditional image segmentation algorithms are threshold method,graph cut method,Snake model and active shape model(ASM).By implementing the above algorithm,comparing the segmentation accuracy and visualization results,the advantages and disadvantages of traditional image segmentation algorithm in segmenting lung tissue are summarized,and the algorithm is tried to improve.(2)An improved deep learning method for FCN segmentation of lung tissue was proposed.The operation of replacing the fully connected layer with the convolution layer realizes the transition from CNN image level classification to FCN pixel level classification.In the improved network structure,the problem of loss oscillation during training is solved by adding L2 regularization;the method of removing black pixel blocks in the image by data clipping solves the false positive problem of segmentation result and improves the segmentation accuracy;In order to obtain better segmentation details,the 32x upsampling and 16x upsampling models were used as pre-training models to obtain the final 8x upsampling model.By improving the training method of the model,the MIoU of the segmentation target can be increased to 90.5%.(3)An improved deep learning method for segmentation of lung tissue in Deeplabv3+network is proposed.In this method,the traditional convolution is replaced by a cavity convolution that can obtain a larger receptive field range,and the semantic feature information and multi-scale feature information are fused on the basis of the FCN sampling the semantic feature information.In the upsampling structure,the Xception network not only improves the boundary information,but also optimizes memory efficiency.By adjusting the crop’s crop size and sample rate combination,95.3%of MIoU was achieved on the public dataset.