Research On Multimodality Medical Image Analysis Methods Based On Deep Convolutional Neural Network

As the population ages,there is an increasing number of the patients suffering from cardiovascular disease,such as cerebral thrombosis,atherosclerosis and aneurysms.Thoracic aortic dissection(TAD)is one of the most dangerous and complex cardiovascular diseases,and the mortality rate is very high(approximately 80%).Hence,accurate segmentation of TAD is urgently needed for the shape quantification and early rupture prediction of thoracic aortic aneurysm,the outcome can provide effective geometric evidence for personalized treatment plan formulation.Additionally,except for degradation and damage of aortic wall factors causing thoracic aortic aneurysm rupture,another key factor is intravascular abnormal hemodynamics,i.e.abnormal changes in blood flow velocity,pressure,wall shear stress and other indices.However,in sickle cell disease(SCD)blood,the heterogeneity morphology of red blood cells(RBCs)has close relathionship with the alterations of the aforementioned intra arterial bio-rheological and hemodynamics properties as a result of the heterogeneity of SCD RBC shapes.Thus,in order to better understanding the pathological mechanism of sickle cell anemia from microscale level,it is important to develop an automated,high-throughput and ex-vivo classification method for the heterogeneity RBC patterns.In addition,to investigate an efficient RBC morphological parameter quantification method can help a more in depth and multiscale morphological assessment for SCD.Additionally,with various medical imaging technologies emerged gradually,due to the high spatial resolution and strong penetrating power characteristics,optical coherence tomography(OCT)and ultrasound(IVUS)imaging technologies have become two major imaging means for the intravascular lesions(e.g.plaques,calcification,dissection,and thrombosis,etc.).And microscopy imaging technology is a useful tool for visualizing RBC morphology in the blood.However,the increasing large number of medical imaging data brings about excessive burden for physicians to read images using their naked eyes.Moreover,artificial judgement is very time-consuming and full of subjectivity.Therefore,studies on multi-scale medical imaging analysis methods for TAD segmentation and SCD RBC pattern classification can lead to better visualization of the macroscale TAD geometric structure and microscale RBC type prediction,hereby provide necessary patient-specific parameters for the blood flow dynamics simulation.Thus,the thesis mainly focus on multiscale medical image analysis methods for thoracic aortic dissection and sickle cell anemia,OCT image enhancement,thoracic aortic dissection segmentation,SCD RBC pattern prediction and classification and cell morphological parameter calculation and quantitative analysis.The contributions of this dissertation are as follows:(1)In order to figure out the low contrast and unclear boundaries around blood vessel in OCT images caused by speckle noise and intrinsic signal attenuation and low penetration ability,an OCT image enhancement method based on variable-order fractional differential operators is proposed.The algorithm adopts three different fractional discretization formulas,and generate 5 different discrete coefficients sets under different discretization orders.A directional symmetric fractional differential operator(or mask)is constructed by the discrete coefficients.In order to further improve the performance of traditional fractional differential operator with same order for the weak edge enhancement issue,and a fractal dimension estimation method is applied to extract the features of image texture roughness,then realize the image contained in the area of different texture information by change order of the fractional differential operator.The experimental results show that the proposed method can effectively reduce the speckle noise in OCT images while enhancing the boudnary of vessel wall in the OCT image.It can conquer the weakness of fractional differential operator with same order enhancing the edge while enhancing high frequency noise as well.Furthermore,the enhancement comparison results among different fractional order of different discrete order is analyzed,and the second order backward finite difference method has relatively good enhancement effect.Moreover,the proposed method is also applicable to natural image enhancement.(2)In order to achieve image segmentation of thoracic aortic dissection with complex anatomy structure,a thoracic aortic dissection segmentation method based on multimodality image fusion and deep CNN is proposed,the proposed multimodality image fusion approach can effectively combine both superiorities of ultrasound imaging and OCT imaging for a clear vessel lumen,false lumen and thrombosis under the artery wall.The experimental results show that the trained deep convolutional neural network model in conjunction with preliminary image fusion operation can get a good performance on the false lumen and thrombus segmentation than the deep CNN model directly learned from single OCT dataset,thus it can effectively figure out TAD segmentation discarding the thrombus contour.Through the comparisons with other two existing methods,our proposed method can effectively segment the true lumen,false lumen and thrombosis.Furthermore,the evaluation results of our method with corresponding histology images show that the proposed method has a better performance on the TAD segmentation than the state of the art.(3)In order to realize an automatic,high-throughput,ex-vivo classification method for the RBCs in sickle cell anemia,and improve the traditional cell classification methods that can only classify the RBCs into two categories,normal and abnormal cells.In this thesis,we propose an automated,high-throughput,ex-vivo RBC classification method based on deep CNN.The proposed method combines with an automated RBC extraction algorithm for microscopy images,and applies a RBC patch image normalization method based on linear intensity mapping to generate the normalized single RBC patches.Subsequently,the complex hierarchical patterns of SCD RBCs are extracted through the customized deep convolution neural networks framework,and realize automatic prediction of RBC categories.The experimental results show that the trained two kinds of deep CNN models in coarse labeling and refines labeling partition level can effectively predict 5 and 8 kinds of complex SCD red blood cell types with accuracy 89%and 89%respectively.Compared with the traditional single two kinds of classification methods,the proposed method can effectively distinguish the heterogeneous SCD RBC shapes,and achieve a fast prediction outcome than using the conventional machine learning methods based on handcrafted features extraction operation.(4)To solve the quantitative analysis of the morphological parameters of SCD RBCs,and provide the shape parameters for the cell biomechanical simulation,a multivariate shape parameter calculation and quantitative analysis method is presented for the SCD RBCs.The method adopted an improved random walk model to realize the accurate RBC boundary contour extraction,and conduct the nonlinear elliptic fitting for the obtained red blood cells contours in different shapes.Then,three different types of RBC shape factors are calculated based on the fitted contours.Finally,statistical analysis is performed on the calculated parameters.Experimental results show that the improved random walk segmentation method can effectively detect the boundary of RBC from raw microscopy images,and the calculated RBC shape parameters is very useful for the multi-scale morphological evaluation of SCD RBCs.Additionally,it can help to verify the effectiveness of the aforementioned automatic classification method.In conclusion,the proposed methods can effectively realize the OCT image enhancement and accurately segment the true/false lumen and thrombus from the interlayer of thoracic aorta in TAD.Moreover,the studies can also effectively realize automated,high-throughput,ex-vivo red blood cells classification and shape parameters calculation and quantification.Thus,these studies play a significant role in promoting the medical imaging diagnosis,treatment and prediction level from a multiscale perspective.