Research On Acute Ischemic Stroke CTP Image Segmentation Based On Deep Learning

Acute ischemic stroke is a life-threatening cerebrovascular disease that causes ischemia and neuronal necrosis in the brain,which could lead to a variety of severe neurological dysfunctions.In this disease,timely identification of ischemic areas of the patient’s brain can reduce the risk of disability or death.Computed tomography perfusion(CTP)imaging provides a large amount of information on brain perfusion,which could help determining the type and extent of the stroke.In medical image segmentation,deep learning-based methods render higher accuracy,better transferability,and more adaptive learning ability.In this paper,we propose a deep learning-based CTP image segmentation method for acute ischemic stroke,which achieves automatic CTP image segmentation and can assist physicians locating the lesion area of patients in a more efficient manner.The research work in this paper can be summarized in the following three aspects.1)To address the shortcomings of convolutional neural networks in acquiring insufficient global semantic information and the severe imbalance of stroke image modality data,the Pool-UNet model is proposed and used for CTP image segmentation in acute ischemic stroke.Specifically described,in order to better extract the global features of the feature map.An advanced Transformer-like architecture is used to learn the multiscale information and establish the remote dependencies of feature information;then,to effectively fuse the global features and local feature information,the advantages of playing both Transformer-like and CNN networks are fused.Subsequently,we apply DSE-RES module to focus on the lesion region in the aim of effectively solve the modal class severe imbalance problem.Finally,a CTP medical image segmentation method based on Pool-UNet model is proposed.Experiments on the ISLES-2018 dataset show that the evaluation metrics such as precision,recall,Dice coefficient,and Hausdorff distance achieve 66.82%,56.54%,56.04%,and 21.14 mm,respectively,in order to perform well.2)The FP-UNet model is proposed to address the problem that the data of different modalities of CTP are not able to be deployed effectively;it also addresses the shortcomings of small lesion edge segmentation.A nonlinear dimensionality reduction method for high-dimensional images is applied to make the original CTP images dimensionally reduced in time dimension.The FP-UNet model firstly fuses Unet,Seg Net,Haar wavelet decomposition and Pool-UNet for model fusion,and proposes a four-path encoder-decoder feature fusion model.Then,the model framework is constructed based on the advantages of self-encoder that could improve feature representation of low fractional rate feature maps.And deep supervised learning structure can reduce fine-grained feature loss.The experiments on the ISLES-2018 dataset show that the evaluation metrics such as precision,recall,Dice coefficient,and Hausdorff distance reach 68.43%,64.38%,58.35%,and 20.56 mm,respectively,in that order,which are excellent performances.3)We develop an automatic CTP image segmentation visualization system for stroke.The system was built using the python language and Py Qt5 framework,equipped with the core technology of two deep learning-based CTP image segmentation algorithms for acute ischemic stroke proposed in this paper.It realizes the functions of case picture display,predicting lesion area,and saving lesion results,which could assist doctors in quickly identifying the patient’s stroke lesion area in CTP images,helping doctors to more accurately diagnose and locate the patient’s lesion location,and improve the diagnostic accuracy.