Study On Image Analysis Of Hyper-actue Ischemic Stroke

Ischemic stroke is one of the most common cerebrovascular diseases,which has the characteristics of high morbidity,high mortality,high disability and high recurrence rate,and seriously threatens people’s life and health.One of the most effective treatments for ischemic stroke is thrombolytic intervention in hyper-acute stage,but not all patients are suitable for thrombolytic therapy.With the development of computer technology,computer-aided diagnosis plays a more and more important role in clinical practice.Based on neuroimaging analysis,the personalized thrombolytic therapy for ischemic stroke by computer-aided diagnosis technology is a research hotspot.However,there are still problems in the existing thrombolytic solutions,leading to under-treatment or over-treatment.The key to image analysis of hyper-acute ischemic stroke is precise segmentation of cerebral ischemia-related regions and tissues.Due to the complexity of the problem,the traditional image segmentation methods are difficult to meet the requirements.The emergence of new theory and technology of artificial intelligence provides new ideas for precise segmentation of brain ischemia related regions and tissues.Through analyzing and investigating a large number of relevant references,the image analysis of hyper-acute ischemic stroke has been studied systematically.The main research contents and results are as summarized as follows.The first part of the thesis is about segmentation of hyper-acute ischemic areas based on sparse representation theory.A supervised learning segmentation framework based on sparse representation is proposed to solve the problems of high heterogeneity,ill-conditioned edge and artifact noise in hyper-acute cerebral ischemia.In the training phase,the classification dictionary is generated by combining the sub-class sample dictionaries to improve its discrimination power.In the testing stage,a fast thresholding method is applied to get the region of interest(ROI).Only intrinsic voxels of ROI are classified based on residual error of sparse representation,which improves the computation speed(< 7s)and the segmentation accuracy(Dice coefficient is 0.755±0.118).The second part of the thesis is on recognition of hyper-acute ischemic areas by sparse representation based bag-of-features.Sample imbalance is a common problem in pathological tissue segmentation tasks.In this study,a hyper-acute cerebral ischemic region segmentation framework composed of voxel classification and region recognition is proposed.Firstly,the training samples are acquired by balanced sampling to train a random forest classifier,which is applied to the training data to obtain the initial segmentation of the cerebral ischemia region.Then,the sparse representation based bag-of-features are extracted from regions of initial segmentation and gold standard segmentation,and applied to train a support vector machine classification model for cerebral ischemia region recognition.Experimental results show that the segmentation accuracy of the proposed method is better than that of random forest segmentation(0.774±0.117 vs.0.606±0.211).The last part of the thesis is on segmentation of veins from susceptibility weighted image based on convolutional neural network.The gray distribution of veins highly overlaps with that of other tissues.Veins of ischemic stroke exhibit high inter-and intra-appearance variability due to ischemic severity.In addition,the slender shape of veins aggregates the segmentation difficulty.We propose a fully-convolutional network structure to segment veins.With the help of dense connection,multi-scale features are fused and reused for voxel label prediction.In order to eliminate the influence of sample imbalance and avoid the smooth effect of Dice loss function on small veins,a weighted hybrid loss function is used during training to further improve the segmentation performance.Experiments show that the proposed method could yield a higher Dice coefficients(0.756±0.043)than other 3 relevant methods.