Research On Pneumothorax Segmentation Method Based On Full Convolutional Neural Network

Among the common chest diseases,pneumothorax is the only pulmonary emergency,and it is very important to get a diagnosis in time.In addition,people are paying more and more attention to physical examination,and the number of diagnostic media X-ray images is increasing,and the X-ray chest X-ray image itself is very complex,its tissue contrast is low,and the boundary between organs and tissues is irregular.At the same time,the regional characteristics of pneumothorax lesions are not obvious,the diagnosis depends heavily on the experience of radiologists,and misdiagnosis and misdiagnosis often occur.In addition,the number of patients is much larger than the number of doctors,causing radiologists to work under pressure and lack of energy,which is more likely to affect the diagnosis of diseases.In addition,traditional segmentation algorithms often require manual intervention and cannot achieve fully automated segmentation of the lesion area.Therefore,efficient and accurate computer-aided tools are urgently needed.We compares the pneumothorax segmentation results of different fully convolutional neural network segmentation models,selects the best baseline model of the segmentation algorithm,and proposes an improved U-Net network model segmentation algorithm based on this to realize the automatic segmentation of the pneumothorax region on the chest radiograph.The network is based on the codec architecture.The structure in the U-Net encoder is replaced with the Res Net50 structure,and the residual learning module is introduced to improve the feature learning ability.The decoder uses deconvolution to complete upsampling to restore the feature map to the original image size.The feature fusion method of the U-Net structure is still used between the encoder and the decoder.In the training phase,in order to speed up the network convergence,the Res Net50 pre-training parameters are called and the fine-tuning operation is performed.Secondly,a multi-loss function training method is proposed,which combines the Dice loss function with the boundary-based loss function(Boundary Loss),which improves the segmentation performance of the model to a certain extent.Finally,the segmentation results predicted by the model are optimized using the fully connected conditional random field to correct the misdivision of pixels predicted by the network.Traditional segmentation algorithms often require manual intervention and cannot achieve automatic segmentation of the lesion area.To solve the above problems,compare the results of pneumothorax segmentation achieved by different fully convolutional neural network segmentation models,and propose an improved U-Net network model algorithm to achieve automatic segmentation of pneumothorax.This network maintains the codec architecture and uses the U-Net encoder The structure in is replaced with the Res Net structure,the residual learning module is introduced to improve the feature learning ability,the decoder uses deconvolution and upsampling to restore the feature map to the original image size,and the encoder-decoder still uses the U-Net structure Feature fusion method splicing,the training stage calls Res Net50 pre-training parameters to speed up the network convergence,and then proposes a multi-loss function training method,which uses the method of combining the Dice loss function and the boundary-based loss function Boundary Loss,which improves to a certain extent Model segmentation performance.Finally,the segmentation results predicted by the model are optimized using the fully connected conditional random field to correct the misdivision of pixels predicted by the network.The improved U-Net network model pneumothorax segmentation method in this paper realizes the pixel-level automatic segmentation of pneumothorax.Compared with the segmentation results predicted by the network model,the optimized segmentation results are closer to the real annotations.In the final segmentation performance,the Dice similarity coefficient is stable at 0.851,the Jaccard coefficient is stable at 0.769,and the pneumothorax pixel classification accuracy rate reaches 0.879.The experimental results show that the method in this paper has a good effect on the task of pneumothorax segmentation.