Research On Ischemic Stroke Lesion Segmentation Based On Deep Learning

Acute ischemic stroke,recognized as a popular cerebral vascular disease,has the features of high morbidity and mortality.Accurate diagnosis and timely treatment can effectively restore the blood supply of the ischemic area and reduce the risk of disability or even death.At present,the diagnosis of acute ischemic stroke often depends on the analysis of stroke magnetic resonance imaging(MRI)by doctors,while the therapeutic schedules are usually determined by observing the location,shape,boundary and size of the lesion.But this is time-consuming and laborious,and will introduce subjective differences among observers.It is necessary to establish an automatic ischemic stroke segmentation model,which can help doctors formulate diagnosis and treatment plans quickly and accurately to improve diagnosis accuracy and shorten rescue time.In this dissertation,four automatic segmentation methods of ischemic stroke lesion based on deep neural network are proposed and applied on Ischemic Stroke Lesion Segmentation(ISLES)2015 MR database.The experimental segmentation results are evaluated by the DICE,accuracy,sensitivity and segmentation distance coefficient.Besides,the superior of the proposed scheme in this dissertation is demonstrated by comparison with other exsiting schemes.The main research contents are as follows:(1)An ischemic stroke segmentation algorithm based on two-dimensional(2D)fully convolutional neural network is proposed.Firstly,the algorithm chooses the MRI slices of ischemic stroke and completes the data preprocessing.Secondly,feature extraction of gray level,pixel,location and other information in MRI slices of the lesion is carried out by Ushaped fully convolutional neural network,and rough lesion segmentation is obtained.Finally,the extracted features are subjected to a fully connected conditional random field.By minimizing the energy function of the features,the rough segmentation result is further optimized and the precise segmentation of the lesion is achieved.The algorithm is applied to ISLES 2015 dataset.The results show that the algorithm can achive the location function,but the accuracy of the algorithm needs to be improved.(2)An ischemic stroke segmentation algorithm based on three-dimensional(3D)fully convolutional neural network is proposed.In order to effectively utilize the threedimensional context information of the images and improve the performance of 2D fully convolutional neural network,this dissertation further proposes a segmentation algorithm based on 3D cascaded U-Net.In this algorithm,two U-Net are concatenated together,and three-dimensional MRI with seven modalities of DWI,CBF,CBV,T1 C,T2,Tmax and TTP are put into network,the algorithm effectively utilizes threedimensional context information to extract features from multi-modality 3D MRI patches,thus automatic segmentation of lesion can be achieved.The experimental results show that the three-dimensional fully convolutional neural network can improve the accuracy of lesion segmentation.The segmentation accuracy of training set and test set are 0.92 and 0.79,respectively.Other objective indexes are also significantly improved,which can meet the needs of clinical automatic diagnosis.(3)An asymmetric 3D residual U-shaped network and a cascaded 3D depth residual network are proposed.The coding network uses the residual module,the decoding network uses the convolution layer,and the coding and decoding constitute the asymmetry.The algorithm has good convergence performance and high segmentation accuracy.On this basis,a cascaded 3D depth residual network is proposed by cascading an asymmetric 3D residual U-Net with a 3D U-Net.Compared with the previous three-dimensional cascaded U-Net algorithm,the problem of network degradation is well overcome while the accuracy of segmentation is guaranteed.In summary,four automatic segmentation methods of ischemic stroke lesion based on deep neural network are proposed in this dissertation.Inputting 2D and 3D multimodality MRI into various improved networks based on U-Net.The feature information of lesion and non-lesion is extracted to achieve segmentation.In this dissertation,the methods are applied on the MRI dataset,the objective indexes of the algorithm are analyzed,and the objective performance of the algorithm is compared with that of the existing algorithms.The results show that methods in this dissertation can effectively segment the lesion and has clinical application value.