Research On Lung Image Segmentation Based On Improved U-Net

Lung disease is the most prevalent illness in our country,with a persistently high incidence rate.Chest X-ray examination is an effective method to evaluate lung conditions.Currently,analysis of chest X-ray images mainly relies on human labor,resulting in wasted cost and time.Therefore,it is necessary to employ image segmentation techniques for processing lung images.After years of development in image processing techniques,neural networks are regarded as having the potential to perform medical image-related area detection.Among them,segmentation methods can be used to detect the contour of lung fields.U-Net is a semantic segmentation model widely used in medical image segmentation in recent years,but with the advancement of medical imaging technology,the shortcomings of this model have also become apparent.Therefore,an optimized U-Net architecture is proposed to address various problems encountered when using the U-Net network for lung parenchymal segmentation.Extensive ablation studies were conducted to find the optimal model architecture and training plan.Additionally,pre-processing operations of the model prior to training were studied,further optimizing the U-Net model,and the expected results were achieved during the experimental phase.The main improvements of this paper are as follows:(1)Data augmentation operations were performed on the original U-Net,specifically adding morphological operations to enhance the images.(2)The improved U-Net was optimized in terms of learning strategy to accelerate the model training speed,alleviate overfitting caused by model parameter issues,and prevent the network from failing to converge;The encoder-decoder structure of U-Net was also optimized,allowing the network to extract multi-level features of the regions of interest while effectively extracting coarse-grained and fine-grained features of the global semantic scale.This optimization improved the problem of uneven edge pixels in segmentation results and enhanced the quality of semantic segmentation of medical images.(3)Adopting a hybrid loss function to train the network and optimizing the gradient descent method utilized in the original U-Net model.Based on the analysis of improved models in other relevant literature,this paper conducts multiple comparative experiments between the proposed improved model,the original U-Net,and other improved models.Ultimately,it is concluded that compared to other models,the proposed method in this paper can achieve superior segmentation results while ensuring computational efficiency.