Medical Image Segmentation Based On Cyclic-attentional Neural Network

Medical image segmentation is an important basis for clinical diagnosis and surgical assistant diagnosis.However,medical images’ s weaknesses including uneven grayscale,low contrast,and blurred edges of adjacent organs will undermine the segmentation accuracy.Furthermore,the adjacent tissue regions are prone to show pixel overlap.Meanwhile,the shape information of organ vary with different patients and slices which will decline the accuracy of segmentation too.Hence,it is essential to explore the above two problems.In this thesis,CT liver image and MRI brain image have been listed as research object because they share some common segmentation challenges including pixel overlap and deformation.Therefore,the paper proposes a medical image segmentation framework cyclic-attentional neural network,which is utilized to design two different technical solutions respectively.The specific research contents and main innovations are as follows:1.The boundaries of liver region to CT image are not obvious and the liver shape information vary with the different patients and time,so that the traditional image segmentation technology and deep learning segmentation method are still difficult to obtain accurate segmentation results.In view of the above limitation,this thesis proposes a CT liver segmentation method by combining bi-directional convolutional long short-term memory(Bi Conv LSTM)and shape prior knowledge.Firstly,U-Net is utilized to act as the backbone network and then Bi Conv LSTM and Attention Gate(AG)has been integrated as the feature-merged module,which will alleviate the influence of pixel overlap on segmentation accuracy.Secondly,the shape prior pattern of the liver is fused into the main segmentation network by Shape-Net,which can improve the robustness to the deformation challenge of liver images effectively.Finally,the liver segmentation accuracy can be further improved by the improved the active contour loss function.This method has been compared on Li TS and 3DIRCADb datasets by seven quantitative indexes.The experimental results show that the proposed algorithm is effective.2.Due to the blurred boundaries between brain structures and the lack of spatial information between different MRI brain images,the accuracy of MRI brain image segmentation is still unsatisfactory.This thesis proposes a MRI brain segmentation method based on HCBAM and Bi Conv LSTM.Firstly,VGG16 works as the backbone network,and the hierarchical convolutional block attention module(HCBAM)is embedded in the encoding stage,which can effectively alleviate the false segmentation caused by pixel overlap in brain image.Secondly,Bi Conv LSTM module is utilized to obtain the context information of brain image sequence in the decoding stage,and then the accuracy of brain image segmentation can be further improved.The method is compared and analyzed on adult brain image dataset(2013-MRBrain S)and infant brain image dataset(2017-i Seg and2019-i Seg).The experimental results show the effectiveness of the method in the segmentation of brain image.