A Medical Image Segmentation Network Based On Improved U-Net

Since the advent of digital medical image imaging equipment,medical images have played a vital role in medical treatment and diagnosis,and their analysis has become the top priority in clinical treatment.In medical image analysis,accurate segmentation of medical images plays a crucial role in its analysis.In recent years,deep convolutional neural networks,especially U-shaped codecs,have shown strong performance in the field of medical image segmentation.However,due to the inherent challenges of medical images,such as dataset irregularities and the presence of outliers,traditional segmentation methods have not demonstrated sufficiently accurate and reliable results in the field of clinical applications.Based on this,this paper proposes a novel deep learning network Attention BConv LSTM U-Net with Redesigned Inception(IBA-U-Net)for medical image segmentation.First,in order to fuse features of different scales and use parallel convolution kernels of different sizes to extract multi-scale details,this paper proposes an Inception module that is more suitable for medical image segmentation to increase the receptive field of the output feature map and improve the grid.The ability to extract features,while reducing the overall network parameters of the model and improving the GPU computing power,this module can improve the segmentation accuracy and robustness of the network.Secondly,in view of the problem that too much noise in medical images interferes with the segmentation accuracy,this paper integrates the BConv LSTM block and the Attention block and proposes the Bi-Conv LSTM Attention block,which can reduce the semantic gap between the encoder and decoder feature maps and reduce noise.interference problems.Bi-Conv LSTM can dig out the hidden information in the encoder and decoder while retaining more features.The Attention block improves the sensitivity of the model.At the same time,the block is concentrated in the part with special meaning,which can reduce the proportion of wrong segmentation.Thirdly,based on the U-shaped encoder-decoder structure,this paper integrates the Redesign Inception and Bi-Conv LSTM Attention mechanisms of the multi-scale feature fusion characteristics of the deep convolutional neural network Attention BConv LSTM U-Net with Redesigned Inception(IBA-U-Net).At the same time,this paper proposes a loss function more suitable for medical image segmentation based on Dice loss loss function and weighted cross entropy.Finally,the architecture IBA-U-Net proposed in this paper has been tested with U-Net and state-of-the-art segmentation methods on three publicly available datasets(lung image segmentation dataset,skin lesion image dataset,and retinal vessel images).In comparison,each dataset has its unique challenges,and in the experimental part we find that the proposed IBA-U-Net improves the prediction performance.The IBA-U-Net network not only has a slightly lower computational cost,but also has fewer network parameters.The network can effectively and accurately complete medical image segmentation for different tasks with only 45% of the U-Net parameters.