Image Segmentation Based On CapsNet And Its Application In Medical Image

Since the advent of convolutional neural network,it has brought many conveniences to our life.Whether it is the detection of vehicles and pedestrians on the road or the detection and segmentation of tumors in medicine,we can see its application products.However,convolution neural network has some shortcomings.To a certain extent,the proposed vectorization of capsule network has improved it.The vector representation of capsules can contain more detailed information,which is very effective for tasks such as medical image segmentation that need high precision.Based on the capsule network,this dissertation first changes the working principle of the capsule network,and then studies the capsule segmentation network based on the SegCap network.The specific work is as follows:First,the reconstruction loss of capsule network is mean square error.Considering that this is only the corresponding relationship between points,we propose to use different similarity function as the reconstruction loss.When cosine similarity function and structure similarity function are used as reconstruction loss,they will make the network produce different effects and make the reconstructed image more complete and mellow.Then,for dynamic routing,this dissertation proposes exponential coupling coefficient,which can change the distribution of coupling coefficient and expect better separation effect.Through the experiment,it is found that the recognition accuracy of the network will be improved,when the index is from 0.5 to 1.And the method is also applicable to other data sets.Finally,this dissertation studies the SegCap network based on the previous research.For the segmentation task,we use different segmentation loss functions to experiment.It is found that Dice coefficient pays more attention to the details when it is used as the segmentation loss function.And we change the reconstruction loss to cosine similarity to ensure the continuity of the line.Then the coherent enhancement diffusion model is used to enhance the segmented image to connect the tiny discontinuities on the line,which has better effect in medical image processing.