Localization And Segmentation Of Two Kinds Of Medical Imaging Lesions Based On Variation Model And Supervised Learning

With the rapid development of medical imaging technology,the diagnosis of disease is inseparable from the medical imaging.Prostate cancer and atherosclerotic plaque are two diseases that are seriously health-threatening.The imaging technology plays a crucial role in the disease diagnosis.Magnetic resonance imaging(MRI)is a non-invasive imaging technology commonly used in the diagnosis of prostate cancer,and ultrasound imaging is an indispensable tool for the diagnosis of abnormal intima-media complex(IMC)and plaque.However,it is often time-consuming and laborious to visually obtain the infor-mation of lesion from an image by using the traditional manual method.Therefore,in order to automatically extract prostate cancer,IMC and plaque,many localization and segmentation methods have been proposed.Although these methods have achieved some results,there are limitations.In the variational model,the segmentation result is affected by heavy noise,blurred boundary of lesion,low contrast,intensity inhomogeneity and other factors.Therefore,the key problems in the localization and segmentation of prostate cancer and atherosclerotic plaque are to reduce the dependence on mannual initialization,ful-ly mining useful information from MRIs and ultrasound images,eliminate interference caused by complex background,reduce the complexity of parameter determination and improve the universality of the method.Thus,this paper mainly studies the localization and segmentation of prostate cancer based on multi-modality MRIs and the segmentation of the IMC and atherosclerotic plaque in the carotid ultrasound images.The main contributions of this dissertation are as follows:1)In order to localize and segment the prostate cancer in the multi-modality MRIs,the random forest with Auto-Context model and the Maximum-Entropy-Based Multiple Kernel Fuzzy C-Means Clustering(MEMKFCM)algorithm are proposed.By introducing the random Haar-like feature,more neighbourhood information can be extracted,which enhances the representation of cancer and improves the ability of distinguishing cancer from non-cancer.More auto-context information are introduced in the Auto-Context model,which optimizes the random forest classifier and makes it localize cancer from the entire prostate region automatically.In the cancer segmentation,the maximum-entropy is introduced to regularize the weights of the kernels in the feature space,which solves the problem of the selection of kernels in the Multiple Kernel Fuzzy C-Means Clustering(MKFCM)algorithm.The proposed MEMKFCM algorithm can make rational use of the extracted features to accurately segment the prostate cancer.The effectiveness of the proposed method is validated on the MRIs of 26 patients.By comparing experimental results,it can be concluded that random Haar-like feature is superior to other traditional features in cancer localization and segmentation.In comparison with the manual seg-mentation results by specialists,the proposed localization method is superior to other conventional methods,obtaining a section-based evaluation(SBE)of 87.1%.The pro-posed prostate cancer segmentation method is better than other methods based on FCM and its improved algorithms,and obtains a sensitivity of 86.0%,a specificity of 93.0%,an accuracy of 91.2%,an Dice ratio of 83.0%,a Jaccard index of 71.7%and an AUC of 89.5%.The localization and segmentation results of the proposed method are closer to the Ground Truth.It demonstrates that the proposed method is promising for prostate cancer localization and segmentation,which can be used for guiding prostate biopsy,assisted diagnosis of prostate cancer and targeted therapy for prostate cancer.2)A fully automatic integrated model to segment the IMC in the longitudinal B-mode ultrasound images of common carotid artery is proposed,which contains continuous max-flow(CMF)algorithm,stacked sparse auto-encoder(SSAE)and random forest.Con?sidering the grayscale-based separability in the longitudinal ultrasound image of carotid artery,an automatic extraction method of the region of interest(ROI)of the IMC based on CMF algorithm is proposed,which abandons the process of manual initialization.In order to overcome the diffculties caused by speckle noise and blurred boundary of the IMC,taking the characteristics of the IMC into account,a SSAE-based reconstruction method is proposed so as to enhance the contrast at the edge of IMC.Random forest and Otsu algorithm are used for the automatic segmentation of the IMC.The mean absolute error,Bland-Altman plot and regression analysis graph are respectively used to estimate the difference,consistency and relevance between automatic segmentation and manual segmentation.Compared with the Ground Truth outlined manually by specialists and other results by some other methods,the proposed approach is superior to other clas-sical and existing algorithms,obtaining the mean absolute errors of 0.028±0.016 mm,0.579±0.288 pixels and 0.582±0.341 pixels on the three datasets from three different imaging devices in different imaging centers(totally 228 longitudinal ultrasound images of carotid artery),respectively.The Bland-Altman plots show the higher consistency between the proposed method and manual segmentation.The mean differences between IMT measurements and the true values are 0.005±0.065 mm,0.031±0.647 pixels and 0.011±0.677 pixels.The regression analysis graphs show that there are the high corre-lation between the results obtained by the proposed automatic method and the manual segmentation.The correlation coefficients between the IMT measurements and the true values of the three datasets are 95.2%,93.8%and 99.0%,respectively.It indicates that the proposed method is robust for the data from different imaging devices,and has the superiority in the IMC segmentation from the ultrasound images of carotid artery.3)A model is proposed for the atherosclerotic plaque segmentation in carotid ultra-sound images by combining random forest and correntropy-based level sets(RF+CLS).The simple linear iterative clustering(SLIC)and AdaBoost are used to extract the ROI of blood vessel simply and quickly.A supervised learning algorithm is used to extrac-t the vessel region covering the plaque,so that the plaque segmentation is completely implemented in the vessel region.Random forest is used to pre-segment the plaque and obtain the initial boundary,which can initialize the model automatically.Considering the discontinuous boundary of plaque or intensity inhomogeneity in the ultrasound im-age,a point-based local bias-field-corrected image fitting(LBIF)method is introduced to fit the image,which makes the proposed method more robust.In order to overcome the limitation that level sets model is sensitive to noise,the correlation entropy is used to measure the distance between the image intensity and the mean of intensities,which can adaptively decrease the variation rate as the noise occurs and greatly suppress the heavy noise.The proposed method is validated on the atherosclerotic plaque segmentation in 25 carotid ultrasound images.By comparing the results,it demonstrates that RF+CLS is the most effective for plaque segmentation,obtaining a Dice ratio of 90.6±1.9%,an Jaccard index of 83.6±3.2%.Moreover,by comparing the standard deviation of each index obtained by the proposed method and other existing methods,it can be found that the proposed method is more robust for segmenting the atherosclerotic plaque in carotid ultrasound images.It also shows that the proposed method is more helpful for clinicians to obtain much more useful information of plaque,make accurate diagnosis and establish effective treatment plan in time.In this paper,the random forest with Auto-Context model and MEMKFCM algo-rithm are proposed to automatically localize and segment the prostate cancer in multi-modality MRIs.To automatically segment the IMC in the ultrasound image of common carotid artery,an integrated method is proposed,which contains CMF,SSAE and ran-dom forest.A model combining random forest and correntropy-based level sets is pro-posed,which achieves the automatic segmentation of atherosclerotic plaque in the carotid ultrasound image.