Study On Computable Modeling Method For Abdominal CT Segmentation

Abdomen image segmentation is a hot issue in medical image processing and analysis,and it is also a very challenging topic.Due to its high resolution,Computed Tomography(CT)can display clear anatomical structures and diseased tissues,and is one of the main examination methods for disease diagnosis in clinic.Automated seg-mentation of abdominal organs and tissues based on CT images is an important first step in computer-aided diagnosis systems.It can provide quantitative information for surgical planning,radiotherapy planning,ablation planning abdominal diseases,which has great practical significance.At the same time,image segmentation is closely re-lated with image registration,disease identification&classification and other issues,and thus has important theoretical significance.This paper focuses on three typical abdominal segmentation problems,i.e.liver segmentation,abdominal multi-organ segmentation,and muscle&adipose tissue segmentation.We take advantage of the theory of deep learning,convex optimization,graph theory,and medical image knowl-edge to establish a series of effective computable mathematical segmentation models,as well as utilize corresponding fast algorithms to solve the proposed models.The main work in this paper includes:Firstly,this paper establishes a fully automatic liver segmentation model based on deep convolutional neural network and global optimized surface evolution.There are a series of difficulties in liver segmentation from CT images.In addition to dif-ficulties brought by the complex background,blurred boundary,the great variability of position,shape and size in different individual' livers,livers with disease also ex-hibit great inhomogeneity in intensity and texture characterization.To tackle these problems,we firstly utilize a deep CNN model to automatically detect liver location and learning the liver probability map as prior information.Then,we establish a re-finement segmentation model.This model introduces a new regional item,which can adaptively utilize the global or local prior information to estimate the gray-scale and texture distributions in different liver regions.Finally,we use a convex optimization-based surface evolution algorithm to solve the proposed segmentation model quickly.Our method has high segmentation accuracy,low computation time,and high degree of automation,which can well meet the actual clinical needs.Secondly,this paper proposes a multi-organ joint segmentation method based on three-dimensional full convolutional neural network and Potts model.We first pro-posed a three-dimensional full convolutional neural network for automatic localization and initial segmentation of multiple abdominal organs.In order to refine the initial segmentation boundary,a new multi-region segmentation model with region disjoint constraint is proposed.This model fuses the information of image gray level,gradient,and prior probability.It also utilizes region competition to overcome the difficulty in labeling the border point caused by border ambiguity and region adhesion,and border point attribution.Finally,we use a time-implicit multi-phase level set algorithm to quickly solve the model.Compared with traditional multi-organ segmentation meth-ods,this method does not require registration or model initialization.A large number of data experiments show that the proposed model can efficiently and effectively seg-ment the liver,kidney and spleen,with accuracy reaching the advanced level of current methods.Finally,this paper proposes a fully automatic segmentation pipeline for abdomi-nal muscle and adipose tissue under the framework of multi-atlas label fusion(MALF).The key in the segmentation of muscle and adipose tissue lies in accurately extracting the structure of the abdominal wall.However,the shape of the abdominal wall varies greatly between different axial slices inter-and intra-individual.Another region of in-terest is the spindle-like psoas muscle,which has very small area in two-dimensional image and is very difficult to detect and segment.In order to solve these problems,we first use the slice-wise multi-atlas label fusion to automatically segment the abdominal wall and psoas muscle.Then,propose a segmentation flow with layer-wise 2D multi-atlas label fusion framework.(1)Firstly,the PC A dimension reduction method is used to perform atlas selection;(2)Then,according to the special shapes of regions to seg-ment,the prior probability maps from MALF are integrated into an Augmented Active Shape Model and a three-dimensional deformable model to optimize the segmentation results of the abdominal wall and psoas muscle,respectively.Finally,skeletal mus-cle,visceral fat and subcutaneous fat areas are extracted using the pre-defined HU ranges.The validation on the clinical dataset shows that the segmentation process has high segmentation accuracy and requires no manual intervention,which is potential in clinical application.