Research On Lung CT Image Segmentation Algorithm Based On Feature Squeeze Excitation Network

Lung CT image segmentation is a basic problem in lung medical image processing.Some studies have used deep neural network to achieve segmentation of lung CT image lesions.Pulmonary nodules are variable in shape and size,and the target area is small.In order to improve the accuracy of automatic segmentation of pulmonary nodules,this paper carried out deep learning segmentation method of pulmonary nodules,and proposed a lung CT image segmentation algorithm based on feature compression activation network.For U-Net model in the presence of pulmonary nodule segmentation task output low precision problem and improve degradation problems brought about by the network complexity,this article in the U-Net of encoding and decoding module,path to add the feature compression activation through to channel characteristic figure and recalibrate the directions of the space to capture key characteristics,and inhibit the expression of meaningless characteristics.The feature compression activation module enhances the representation ability of the network model,extracts the features of pulmonary nodules at different levels,and deepens the number of network layers.In order to avoid the problem of gradient disappearing in the deep layer of the network and accelerate the convergence of the model,residual module is introduced.Residual connection simplifies the training process and improves the performance degradation caused by the improvement of model complexity.At the same time,a new loss function was designed aiming at the small area occupied by pulmonary nodules in CT images.The improved loss function integrated classification error and segmentation error to avoid the overfitting problem and make the model training process more stable.Based on the above improvements,this paper designed a lung image segmentation algorithm with a code-decoding structure called SERU-NET,which can achieve end-to-end segmentation of pulmonary nodules.In this paper,the effectiveness of SERU-NET algorithm is verified by several comparative experiments.For the segmentation task of pulmonary nodules,training and testing were carried out on the lung CT image dataset LDC-IDRI.Experimental results showed that the improved model achieved a Dice value of0.640,which was %7.92 higher than the U-NET model.In order to verify the effectiveness of the improved module,a comparative experiment is carried out in this paper.New models are obtained by removing the optimization mechanism in encoding and decoding paths respectively.The segmentation accuracy of these networks is improved compared with U-NET,but still lower than Ser U-NET,which further verifies the rationality and effectiveness of the model proposed in this paper.