Researches On Segmentation Of Lung Nodules Based On Hybrid Operator And Attention Mechanism

Lung nodules are the criteria for early diagnosis of lung cancer.Accurate segmentation of lung nodules is one of the key steps in formulating treatment plans,which can greatly improve the cure rate of lung cancer patients.However,the existing lung nodule segmentation networks have faced the problems of serious feature loss in the process of capturing features,and the local context information is insufficient,resulting in inaccurate segmentation of small nodule region and inaccurate edge segmentation.To solve the above problems,based on the deep learning technology,aiming at the lung nodule segmentation task in CT image,this paper proposes two lung nodule segmentation models to improve the accuracy of lung nodule segmentation.The main contents are as follows:（1）Firstly,the Hybrid Operator combining convolution operator and involution operator is designed to deeply mine the feature representations of channel domain and improve the potential diversity of feature representation.Then an end-to-end lung nodule segmentation model HOU-Net based on Hybrid Operator is proposed.HOU-Net is an encoder-decoder structure based on the U-Net being widely used in medical image segmentation.In the encoder,the Hybrid Operator is constructed to extract the features of the picture.In the decoder,the deconvolution operation is used to combine the features extracted in the corresponding stage with the upper features to generate the final segmentation result graph.Finally,the effectiveness of the Hybrid Operator is verified by experiments,and on the lung nodule segmentation data set,this method is combined with FCN＿32s,Seg Net,R2 UNet,Attention U-Net and other segmentation networks,which proves that the model is lighter and more accurate.When applied to the task of thyroid nodule segmentation,it also shows good generalization.（2）In order to improve the ability of extracting nodule edge information and segmenting nodule fuzzy edge region,a segmentation network GMAUNet based on grouping mixed attention mechanism is designed in this paper.Firstly,this paper designs a "Plug and Play" grouping mixed attention module GMA.By combining channel attention and spatial attention,the input features are grouped to calculate attention,and the feature map is "shuffle",which improves the expression of feature structure information and makes the generated edge feature more accurate.GMA can enhance local correlation features and suppress irrelevant features in space and channel domain,so as to obtain different semantic features more effectively.The experimental results show that GMA has excellent feature information extraction ability.GMAUNet uses GMA module to extract the features of lower sampling layer and transfer them to the corresponding upper sampling layer to complete the restoration of the detailed features of the lesion area.In order to verify the robustness of the model,this paper applies it to the thyroid nodule segmentation task.The results show that the neutral energy of the model is improved in all three indexes.Finally,the ablation experiments of HOU-Net and GMAU-Net are carried out.The results show that the two methods have strong competitiveness in segmentation tasks.（3）To solve the problem of data class imbalance,this paper combines GDL loss function and CE loss function,designs a new hybrid loss function,and uses weight parameters λ assign weights to the CE loss function.In the experiment,this paper first analyzes λ,and then the effects of GDL loss function,CE loss function and hybrid loss function are compared.The results show that the hybrid loss function has the best segmentation effect.