Medical Image Segmentation Based On Deep Learning

Medical image has become more and more important for doctors5 diagnosis.As the number of medical images rapidly grows,medical image analysis based on computer is one of the main researches in the modern medical image science.Browsing many medical images makes the diagnosis process tedious for doctors,and the diagnosis is easily influenced by doctors' experience and fatigue.Computer-aided diagnosis system can help to reduce doctors' burden and improve the accuracy of diagnosis.Automated segmentation of three dimensional(3D)medical images plays one vi-tal role in modern clinical diagnosis and treatment,for example,the segmentation of brain tumor in magnetic resonance images(MRI)is good for the removal surgery and postoperative observation.Conventional image segmentation methods usually require prior knowledge when segmenting the brain tumors,such as the shape and some fea-tures of the brain tumor,which limits the use of conventional methods in practice.The brain tumors with complex and variable shapes can occur anywhere in the brain,which causes that it is difficult to segment brain tumors in 3D medical images.In recent years,deep learning has achieved great success in the field of computer vision,especially the method of segmentation based on convolutional nerual network(CNN)get better result than the conventional methods for the natural light images' segmentation.Deep learn-ing is one data-driven method to extract relevant features automatically.This paper studies the method of automatic segmentation based on deep learn-ing for the segmentation of brain tumors in 3D medical images.The encoder-decoder symmetric architecture serves as the basic framework of the end-to-end segmentation convolutional network.First of all,the brain tumor segmentation networks based on 3D convolution operation are presented.However,there are some problems of the segmen-tation networks based on 3D convolution:the large number of the network parameters,the difficulty of training and the huge calculation.So the pseudo three dimensional con-volution operation is introduced,further one brain tumor segmentation network based on the pseudo three-dimensional convolution is proposed,which is fine-tuned on the two-dimensional pre-trained segmentation network model,then obtains better segmen-tation results.The end-to-end segmentation convolutional networks have small effect on the segmentation of the brain tumors' internal structure,therefore one two-step seg-mentation strategy is proposed:firstly extract the bounding boxs containing the brain tumor,then doing the segmentation of the brain tumors' internal structure from the bounding boxs.Three-dimensional medical images can not only be viewed as a whole,but also a series of two-dimensional slice images.So we implements one recurrent neural net-work combining convolution operations to complete brain tumor segmentation.The recurrent neural network is used to extract one-dimensional temporal features across two-dimensional slice sequences,while the convolution operations extract the two-dimensional spatial features of each slice image.Finally integrate the one-dimensional temporal features and two-dimensional spatial features to complete brain tumors seg-mentation in 3D brain medical images.