Study Of Random Forests Based Medical Image Segmentation Algorithms And Applications

With the development of the medical imaging and computer science,the technology of medical image processing has been greatly improved.This thesis focuses on image segmentation,the core problem of medical imaging processing.Medical image segmentation is an important step for the diagnosis of diseases,quantitative analysis of organs,and 3D reconstruction of human body.For the common difficulties of different image modalities of different organs,the thesis studies the random forests and the combination with model based method.To improve the accuracy,efficiency,and robustness of segmentation,feature design,extraction and selection are also studied.The novel algorithms are described below.(1)For the difficulties of localization of renal cortex and segmentation of kidney components,the automatic segmentation of kidney components based on modified Random Forest and Active Appearance Models is proposed.The proposed method is consisted of two main parts: localization of renal cortex and segmentation of kidney components.In the localization step,we propose a renal cortex localization method by combining 3D Generalized Hough Transform(GHT)and 3D Active Appearance Models.Compared with the traditional AAM,the efficiency of the proposed localization method is greatly improved.In the segmentation,a modified random forests method is applied with many different kinds of features.By the proposed method,each kidney component can be quantitatively analyzed.(2)Based on the method described above,the fast segmentation of kidney components using Random Forests and Ferns is proposed.Although the method based on random forests and AAM can segment the kidney into four parts accurately,it still has some limitations which mainly cause by AAM.In order to improve the efficiency and keep the high accuracy,we propose a more efficient approach by strategically combining random forests and random ferns.In addition,we proposed the Potential Energy Features which can describe the spatial relationship of kidney components and the adjacent organs.The performance of the proposed method has been validated by many different experiments.These experimental results validated the performance of the proposed method.(3)For the difficulties of segmenting OCT images with different diseases,the automatic segmentation of Macular Hole,Cystoid Macular Edema and Retinal layers based on random forests is proposed.We propose an automatic method based on random forests with the novel feature group including Potential Energy Features to segment the disease areas as well as retinal layer structures for abnormal retina.The method is tested on OCT images with CME and MH.The experimental results show the high accuracy and efficiency of the proposed method.