Research On Pneumothorax Segmentation Algorithm Based On Deep Learning

Pneumothorax is a common lung disease,which can cause death in severe cases.X-ray is usually used to check whether the patient has pneumothorax.However,because the X-ray chest film image itself is very complex,the contrast between organs and tissues is low,the boundary shape is irregular,and the boundary of pneumothorax is fuzzy,which makes it difficult for the traditional segmentation algorithm to achieve high accuracy.With the development of deep learning,the image processing methods based on convolutional neural network have been well used in the field of medical image segmentation.Therefore,this paper studies the segmentation algorithm based on deep learning,designs a more powerful feature extractor to solve the problem of blurred boundaries of pneumothorax,and proposes two pneumothorax segmentation algorithms: one is SAWF-SDM model,a dual-task interactive learning pneumothorax segmentation algorithm based on U-Net;The other is MSFF-Trans UNet model,the Trans UNet pneumothorax segmentation algorithm based on multi-scale feature fusion.The former segmentation algorithm can improve the efficiency of learning the boundary of pneumothorax image,while the latter segmentation algorithm can improve the ability of global feature extraction and spatial information location of the model.The work of this paper is divided into the following two parts:(1)The proposed SAWF-SDM model combines abundant boundary information in the Signed Distance Map(SDM),learns binary mask label and SDM label in parallel through multi-task learning strategy,and provides more supervision information for model training through multi-labeling,so that the model can learn more accurate contour information.In addition,in the process of decoding,the boundary information learned by SDM prediction branch is fused into semantic segmentation prediction branch by spatial adaptive fusion,which further strengthens the ability of model feature extraction and improves the segmentation result.(2)Based on the Trans UNet model,the proposed MSFF-Trans UNet model designs and adds a multi-scale feature fusion module in the encoder part,which alleviates the problem that the positioning ability of the Transformer segmentation architecture is limited due to the lack of shallow details.In addition,the fusion of the jump connection part is also improved,which can effectively narrow the semantic gap between the corresponding levels in the coding and decoding process,and alleviate the influence of image noise caused by the fusion of features with semantic gap on the segmentation results.