Study Of Lung Parenchyma Segmentation Based On Data Enhancement And U-Net Model Improvement

The most common means of clinical diagnosis of lung disease is through the observation of computed tomography(CT)images of the patient’s lungs.With the increase in the number of patients with lung diseases in recent years,Computer Aided Diagnosis(CAD)has been widely used for screening and initial diagnosis of lung diseases by processing and analyzing medical data quickly and accurately.Among them,the ability to achieve accurate segmentation of lung parenchyma on CT images is crucial for the initial screening and diagnosis of lung diseases.Traditional segmentation methods not only require manual setting of parameters,but also tend to cause under-segmentation.Deep learning-based segmentation methods not only learn useful features automatically,but also do not require manual adjustment of parameters through a priori knowledge,thus replacing the manual feature extraction-based methods and gradually becoming the primary method in the field of medical image segmentation.However,deep learning-based lung parenchyma segmentation has several challenges and difficulties in practice,such as difficult to obtain datasets and unstable quality of lung CT images.In addition,category imbalance,multi-scale problem and model interpretation problem are also difficult problems that plague lung parenchyma segmentation.In response to these issues,the main research content of this article includes:(1)A data augmentation-based approach for lung parenchyma segmentation.The original data is augmented by generative adversarial networks to address the problems of small size,uneven quality and unbalanced categories in medical image datasets.A generative model ResAttention GAN(RA-GAN)based on conditional generative adversarial network is proposed to solve the problems of unstable network image generation with blurring and inconspicuous boundary distinction.Firstly,the residual blocks are used instead of convolutional layers after the down-sampling of the generator to obtain deeper feature information;secondly,Convolutional Block Attention Module(CBAM)is introduced in the long-jump connection stage to weight the feature map features adaptively to capture richer contextual information;finally,a discriminator based on the generator is used to guide the generator to generate Finally,a discriminator based on the CT image is used to guide the generator to generate higher quality CT images.The effectiveness of RA-GAN-generated images for improving the accuracy of lung parenchyma segmentation is further validated in the segmentation experiments,and data augmentation by RA-GAN provides more high-quality data samples for subsequent segmentation tasks.(2)A lung substance segmentation method based on a double U-shaped segmentation network.For the problems of insufficient feature fusion and serious information loss in the U-Net network and its variants,a double "U-shaped segmentation network is proposed,which is divided into an encoding network and a segmentation network.Firstly,the convolutional layer is added after each downsampling to enable the network to acquire more local information of the image;secondly,the long and short jump connection paths are designed in the segmentation network for multi-scale feature information fusion to solve the semantic gap problem;finally,the bottleneck supervision is applied to the whole network,which can make full use of the features of each layer to avoid information loss and overfitting and improve the generalization ability of the model.Experimental results on the LIDC-IDRI dataset show that the improved U-Net segmentation network segmentation performance based on bottleneck feature supervision is excellent and can be well applied to the segmentation task of lung substance.