Research On Lightweight U-Net Brain Tumor Segmentation Based On Attention Mechanism

The brain tumor is the second most fatal malignant tumor among adolescents in the world.The accurate segmentation of its MRI images is very important for diagnosis and treatment.With the development of artificial intelligence and the improvement of computing speed,deep learning has shown great advantages in biomedical image segmentation.More and more computer-assisted medical technologies are used in clinical medicine.As the cornerstone of medical image segmentation,U-Net has achieved good results in segmentation tasks such as brain tumors,organs,and cell nuclei with the encoder-decoder end-to-end structure,and has provided a basis for subsequent work.However,it is precise because of this structure that the network has a huge amount of parameters,which brings great challenges to the segmentation speed,memory and energy consumption.In response to the above problems,this thesis from the perspective of lightweight models,mainly did the following work:(1)Firstly,the class imbalance is common in image segmentation tasks.Especially in medical images,the problem is more prominent.The dataset used in this thesis also has the problem of class imbalance.To reduce the impact of class imbalance on the accuracy of segmentation,this thesis proposes a loss function named BCEDice Loss for image segmentation.It combines the balance cross entropy loss function based on cross entropy and the Dice loss function based on the Intersection-over-Union through weighted summation.We conduct experiments on the Bra TS dataset and compare the loss function proposed with the balance cross entropy loss function.Better results are achieved in both U-Net and networks proposed,and the class imbalance problem is improved.(2)Secondly,in the network with an encoder-decoder structure,the decoder often only provides a coding vector to the decoder.During the decoding process,the input information almost disappears,and the loss of detailed features is serious.Therefore,this thesis proposes a global attention module,which can better capture long-distance dependence.And three convolution methods of traditional convolution,atrous convolution,and transposed convolution are integrated into the global attention module to calculate the correlation between any two positions in the feature map.Besides,the global attention module combines the spatial attention mechanism and the channel attention mechanism at the same time,and it is verified through experiments that the cascading connection mode inside the module is more conducive to the improvement of model performance.The global attention module can obtain the dependence between the features in all directions from the space and the channel,and encode the important features in the space and the channel through the global attention module,so that the network’s feature representation ability is stronger.(3)U-Net network is considered to be the cornerstone of automatic medical image segmentation.However,due to the huge amount of parameters,problems such as memory occupation,slow speed,and energy consumption often occur.Inspired by the Squeeze Net,this thesis proposes the lightweight network Squeeze U-Net(SU-Net).Use the Fire module to replace the convolutional layer in U-Net,and compress the model parameters by changing the convolution method.In order to ensure the segmentation ability of the network while compressing the parameters,the sampling modules are designed to replace the simple pooling layer and upsampling layer to avoid the feature loss caused by direct sampling and use the loss function BCEDice Loss to train the network.Using U-Net as the baseline network,our network model parameters are compressed to 50% of the original,and the average DSC is increased from 80.40 to 81.34.(4)Finally,combining the BCEDice Loss,global attention module and lightweight model strategy proposed above,this thesis proposes a global attention U-Net named GAU-Net.The global attention module is integrated into the U-Net network,and a large number of residual connections are used in the network.A Fire Module and a designed sampling module are used as the backbone.A simple bypass can be connected to realize the up and down sampling blocks,and its implementation is equivalent in U-Net,there are two layers of convolution and one layer of sampling,but it greatly reduces the number of network parameters.We conduct experiments on the dataset Bra TS2018 and the mean Io U is increased from 0.65 to 0.75 with a parameter amount equivalent to 5.4% of U-Net.At the same time,the reasoning time is also significantly shortened with relatively good performance.The algorithm proposed in this thesis carefully considers the practical problems encountered in the process of algorithm implementation,such as memory occupation,serious time-consuming,difficult hardware embedding,etc.Based on the U-Net,the image segmentation network has been researched and improved,and good results have been obtained in brain tumor segmentation,which provides a certain idea for the future research and development of image segmentation algorithms.