Medical Image Segmentation Based On Deep Learning

Because of the increasing number of medical images,the limited energy of doctors,and even the dependence that some observations of medical images need to rely on those doctors that have complete domain knowledge and rich experience,there might occurs the situation that the diagnostic error would result from the lacks in experience or the fatigue of doctors.The main research of this passage is based on the automatic segmentation of medical images,only computers’ operations on the input image to obtain the segmentation result of the input image without artificial interference.The segmentation results of the traditional image segmentation methods are influenced by the noise easily.And with the development of deep learning,more and more methods about deep learning has been used in many kinds of computer vision tasks.Compared with traditional segmentation methods,using deep learning in semantic segmentation of medical images is less easily influenced by noise.This paper firstly studies the traditional semantic segmentation methods and semantic segmentation algorithms based on deep learning,introduces the principle and then completes the following work:1)As far as the blurred boundary problem in semantic segmentation of the medical images,this paper proposes a semantic segmentation method that is based on boundary information.Because the Octave convolution can not only divide the tensors of feature into two different kinds of information that belongs to different frequency,but also make the information in the low frequency band more compact,reducing the spatial redundancy of low frequency information,this paper proposes decoding the different frequency band information that is decomposed by Octave convolution,getting the segmentation results of the profile information and the detailed information in image,and constraining the segmentation result of the profile information in image by the segmentation result of the detailed information in image at last.Otherwise,this paper also considers the problem of class imbalance in the image segmentation,providing higher weights for rare classes to make segmentation results more accurate by setting the weights in cross-entropy loss function based on the rate that each class accounts for.The following experiments prove that this method called Oct-UNet is more effective than the segmentation method(U-Net)that it depends on,and the number of required parameters is much less than U-Net.2)As far as the connection relations problem within each feature map in Oct-UNet,this paper proposes a method that applies dense connection to the Oct-UNet.By adopting the form of dense connection,the output of a certain network layer is not only related to the input of this layer but also related to the output of the network layer before this layer.This method is conducive to the flow of feature information in the network and can further improve the performance of the network segmentation.The following experiments prove the effectiveness of this method for the segmentation problem.In addition,through the deep supervision,it is found that in order to obtain the accurate segmentation results,more times of the down-sampling operations does not necessarily mean better.This paper is based on U-Net,proposes Oct-UNet from the blurred boundary problem and then adopts dense connection in Oct-UNet to improve the flow of the feature information.These segmentation experiments conducted on the ultrasonic dataset of standard sections of fetal brain verify the effects of these two methods.