Research And Implementation Of X-ray Pneumothorax Segmentation Based On Cascaded Convolutional Network

Pneumothorax is a common respiratory emergency,which occurs when air infiltrates the pleural cavity between the lung and chest wall.Mild cases may cause chest pain,while severe cases can be life-threatening,with a high recurrence rate.Therefore,timely detection and clinical treatment of pneumothorax is crucial for patients.Pneumothorax can be detected through imaging examinations,and compared to other imaging techniques,X-ray examination is cost-effective,fast,and has low side effects.It can clearly show the pathological features and range of pneumothorax and has a good resolution for chest tissues.Its images contain a wealth of information.The pneumothorax area in X-ray images is difficult to identify,and the diagnosis results are largely subject to the doctor’s subjective influence.It is easy to be misdiagnosed or mistaken for other diseases,and small pneumothorax can also be difficult to detect.For doctors,there are often a large number of X-ray images,and they need to spend a lot of time identifying pneumothorax,leading to long working hours and insufficient energy,which may affect the diagnosis of the disease.Therefore,there is an urgent need to achieve automatic examination and segmentation of X-ray pneumothorax.With the development of deep learning,there have been many studies on segmentation methods for X-ray pneumothorax.However,these methods have a series of problems in pneumothorax segmentation,such as over-segmentation,under-segmentation,mis-segmentation,and low segmentation accuracy,which cannot be truly applied in clinical practice.In response to these issues,this article conducts a series of research on the segmentation of X-ray chest images,and the main work carried out is as follows:(1)X-ray pneumothorax segmentation method based on improved UNet++This method is based on the UNet++ network architecture and introduces the Res Ne St module to replace the original VGG16 of the UNet++ network.The Res Ne St model is an improved scheme based on the Res Net model,integrating the multi-path ideas of Google Net and the attention ideas of SENet,and introducing the Split-Attention module to enhance feature extraction capabilities.During decoding,to further improve segmentation performance,we embed the sc SE module in the decoder module of the model,integrating the sc SE module into the decoder to adaptively recalibrate feature maps,suppressing useless features while enhancing useful features without adding too many parameters.We conducted experimental verification on a chest dataset obtained from the clinical report database of Mianyang Central Hospital.The DSC,MIo U,recall,and accuracy of our proposed pneumothorax segmentation network model reached 88.56%,83.46%,88.12%,and 99.44%,respectively,which are superior to other commonly used medical image segmentation networks.The experimental results show that this method can effectively improve the accuracy of X-ray pneumothorax segmentation.(2)X-ray pneumothorax segmentation method based on cascaded convolutional networkTo further improve the segmentation accuracy of X-ray pneumothorax and assist doctors in more accurately calculating the size and quantification of pneumothorax,we propose a pneumothorax segmentation method and pneumothorax ROI extraction method based on cascaded networks.Specifically,the first step is to use the method proposed in the previous chapter to segment the entire chest X-ray image,segmenting out the approximate area of pneumothorax for localization.In the second step,we use the segmentation results from the first step and extract the corresponding original image ROI area for input into the second model for precise segmentation.We then scale the second segmentation result to the size of the first segmentation and replace the ROI area of the first segmentation to obtain the final segmentation result.The DSC,MIo U,recall,and accuracy of the cascaded segmentation network model proposed in this chapter reached 88.72%,83.54%,87.76%,and99.44%,respectively.Compared with UNet,the DSC increased by 3.13%,and the MIo U increased by 2.82%,significantly improving segmentation accuracy.Compared to the method proposed in the previous chapter,the DSC increased by 1.6%,and the MIo U increased by 0.8%.(3)Web-based X-ray pneumothorax segmentation systemCurrently,manual X-ray pneumothorax segmentation is time-consuming and laborious,making it difficult to implement in clinical practice.To address this issue,we combined theory and practical application to design and develop a web-based X-ray pneumothorax segmentation system.This system is developed using PHP and Python and includes user management,pneumothorax segmentation,and pneumothorax display modules.The system is easy to operate,has a quick response time,and can promptly segment the pneumothorax area,helping doctors analyze and diagnose pneumothorax in X-ray images,assisting doctors in quantifying pneumothorax,and improving diagnostic efficiency.