Research On Medical Image Segmentation Algorithm And Application Based On Deep Learning

The first step in the field of medical vision is to determine the internal region of interest,which is one of the key steps of medical image segmentation.Aiming at the task of medical image segmentation,this paper studies the algorithm based on deep learning,and carries out experiments on the nuclear,liver and lung image segmentation data sets to improve the accuracy of medical image segmentation.The main research contents include:Firstly,a nuclear image segmentation algorithm based on convolutional neural network and codec structure(Att Inc Net)is improved.Based on the U-Net framework,the encoder part integrates the channel attention SE and stripe pooling.A lightweight multi-scale module is designed on the cascaded path to reduce the feature difference between encoder and decoder.The decoder part adds a side output structure which can provide multi-level information,and uses ladder and loss supervision training.Experiments on the nuclear image segmentation data set show that Att Inc Net has better segmentation effect than other networks,but this method does not consider the long-distance dependence of modeling.Then,a hybrid two-way liver image segmentation algorithm based on CNNTransformer(Hct Net)is designed.Although convolutional neural network can capture local features,it can not establish explicit long-distance dependence in the global feature space.In view of the limitations of Att Inc Net,Hct Net takes into account the advantages of convolution and transformer,takes the original U-Net as a convolution link to extract local information,and establishes an independent transformer link to capture the global connection and effectively integrate the two at the bottleneck.Experiments on liver image segmentation data set show that Hct Net is superior to other advanced networks,which provides a concise and general basic framework for subsequent research.Finally,a lung image segmentation algorithm based on generated adversarial network and Transformer(Sct GAN)is improved.Most segmentation algorithms use pixel loss monitoring model to learn local information and remote contact,which is difficult to further improve the performance of medical image segmentation.Sct GAN is based on Pix2 Pix framework,the generator uses Hct Net and uses its powerful feature extraction ability to enhance the performance of the generated network,the discriminator part designs a network that convolutes first and then transforms and outputs multi-level results.It can effectively improve the discrimination ability of the discriminator network by evaluating it in different degrees of vision.Experiments on lung image segmentation data set show that Sct GAN has achieved good results.