Research On Related Technologies Of Medical Image Segmentation Based On UNet CNN

Medical images provide intuitive analysis basis for biological research and clinical medical application,and their role are becoming more and more obvious.The objective of medical image segmentation is to describe the boundary of the target area to achieve the quantitative measurement of the image pixel by pixel,according to the characteristics of medical image.It plays an important role in biological tissue recognition,disease evaluation and surgical guidance.At present,deep learning algorithm based on Convolutional Neural Networks(CNN)is one of the most effective methods for medical image segmentation.The U-shaped encoding and decoding structure is the most widely used among many CNN models.However,due to the low resolution,noise interference,partial volume effect and the three-dimensional complexity of human body structure,it is a very challenging research topic to achieve accurate segmentation in polymorphic medical images.In addition,small categories of medical images still receive little attention.Therefore,focusing on the above problems,this paper studies the medical image segmentation algorithm based on Multiphoton Microscopy(MPM)self-built dataset and public datasets under the U-shaped framework.Firstly,a Dense＿UNet model combined with dense bypass mapping is proposed to improve the segmentation accuracy of low-resolution medical images.The feedforward cascade with small convolution structure mode is adopted to deepen the network structure,so that the ability of model detail description is improved and the impact caused by low resolution and inaccurate gray description is reduced.The model can adaptively complete network deepening by modular feature reuse with identity mapping in polymorphic datasets with different resolutions.Then,the modular bypass feature mapping can effectively alleviate the gradient disappearance and control the computational cost.In addition,the MPM in vivo skin cell dataset is constructed in this paper to complete the attempt of CNN segmentation model in this field.After comparing Dense＿UNet with UNet and its variants in self-built and public datasets,the results verify that the dense bypass model can balance the accuracy loss caused by low resolution and uneven gray scale.Secondly,a U-shaped elastomeric UNet(EUNet)is designed to suppress the precision decline caused by noise pollution.The network adopts two ways of horizontal serial fusion and vertical parallel cascade with independent U-shaped paths to enhance the network’s forward convolution mapping capability from the structure rather than within the module,thus breaking the constraints of the original U-shaped structure.Feature abstraction and information fusion are carried out many times to make fully use of the new complementary information to extract pixel features and reduce the pixel loss caused by noise interference.The output layer of EUNet adopts skin-connection closing to identity mapping,which is easier to abstract the subtle fluctuation of bypass feature and improve anti-interference ability.The independent U-path structure of the elastomeric benchmark model provides flexible design redundancy.The test results show that EUNet can effectively resist noise pollution and improve the segmentation performance.Thirdly,in order to overcome the influence of partial volume effect in medical images,spatial and dimensional attention mechanisms with clear mathematical structure interpretation are proposed.Four groups of U-shaped segmentation networks with attention modules are designed by using these two kinds of attention mechanisms.Convolution multiplication operation is performed in spatial attention module to enhance the position information correlation of pixels and the feature of target area,so that the impact of partial volume effect is reduced.Weighted assignment is performed in dimension attention module to complete the adaptive channel correlation response and improve the pixel sensitivity of the target area.In addition,the effective activation function combination is also carried out to improve the nonlinear torsional force and accelerate the convergence ability of the model.The test results demonstrate that the four groups of models have convergent fine-grained segmentation ability.The partial volume effect can be balanced by attention mechanism.To sum up,focusing on the common problems of low resolution,noise interference and partial volume effect in medical images,this paper expands the design vision of Ushaped medical image segmentation from modular feature reuse,structural network design,and attention mechanism with interpretable mathematical structure.