Research On Computer-Aided Diagnosis Of Intracranial Large Vessel Occlusion Based On Deep Learning

Stroke is a type of brain disease that seriously endangers human life and health,and ischemic stroke(IS)is the most common clinical condition.Acute Ischemic Stroke(AIS)caused by intracranial Large Vessel Occlusion(LVO)is the most serious,causing60% of stroke-related disabilities and 90% of stroke-related deaths.However,if patients can be treated effectively within the time of onset,the poor prognosis is greatly reduced,so it is imperative to quickly diagnose the presence of AIS and LVO and to treat the patient appropriately based on symptoms.For the diagnosis of LVO,Computed Tomographic Angiography(CTA)is the main imaging tool.For surgeons,they need to analyze hundreds of CTA slices layer by layer to determine the condition,which is time-consuming and highly dependent on surgeons’ personal work experience.Therefore,it is clinically important to establish a fast and effective computer-aided diagnostic method in order to reduce the time spent by surgeons in viewing CTA images and to speed up the process of completing the diagnosis and receiving appropriate treatment for patients.In order to realize a reliable computer-aided diagnostic method for LVO,this thesis designs two network models for detecting LVO based on the characteristics of the collected data from two different perspectives.One of them is to get the difference information by comparing the symmetrical position of the brain bilaterally based on the balanced number of samples of each category by means of data Amplification,and then explore the difference information to make classification.Another approach is to design a multi-branch structure,which enables each branch to have its own learning focus to alleviate the long-tail problem and thus achieve good classification performance.Finally,for ease of use,a LVO recognition prototype system is developed with the two algorithms proposed in this thesis as the core.The specific work of this thesis is as follows:1.To address the problem of no publicly available study dataset,this thesis constructs the CTA Intracranial Large Vessel Occlusion Dataset(ILVOD)and its augmented version,the ILVOD-Aug dataset.To construct the dataset,the CTA raw image format data was first mapped to image format data by maximum density projection.Then,pre-processing operations were performed by thresholding,maximum mask extraction and affine correction with the help of image processing techniques.Then,images of poor quality and those that did not meet the study criteria were removed.Finally,the image instance-level classification labeling and the related vessel pixel-level segmentation labeling are completed with the help of professional surgeons.In view of the serious uneven number of samples of each type in ILOVD dataset,this thesis constructs another ILVOD-Aug dataset by means of data amplification.For ILOVD-Aug dataset,in addition to the traditional data augmentation methods such as flip,rotation,and elastic deformation,two new data augmentation methods such as manual vessel erasure and image reorganization and replacement were designed to expand the intracranial large vessel occlusion dataset.2.According to the human brain shape structure roughly conforms to the left-right symmetry,A classification network model of large vessel occlusion based on bilateral contrast difference information of the brain was designed.The model obtains the difference information by comparing the corresponding positions of bilateral brains,and then uses the difference information to achieve high-precision recognition results.The inputs are two images of the left and right hemispheres,and then the two hemispheric images are fed into the designed “total-split” structured network to complete the task learning.In addition,we also introduce the deep supervision and comparison loss,which can make the model complete the learning process faster and better by narrowing the spatial distance of the same class and distancing different classes in the feature space.Finally,this thesis explains the decision-making basis of the model with the help of heat map visualization technique,which helps surgeons to judge whether the prediction results of the model are reasonable or not.3.To address the problem of long-tail effect in the ILVOD dataset,We have proposed a Multi-branch Hybrid Network for Large Vessel Occlusion Detection.The network can alleviate the overfitting of the head data and underfitting of the tail data by using the characteristic that the three branches have different learning focuses,thus coping well with the long-tail effect problem.Specifically,The three branches achieve a different focus by using different sampling strategies,The first two branches focus on learning semantic features on images of normal people and patients with vascular occlusion,respectively,while the third branch achieves the final classification decision by weighted fusion of semantic information from the first two branches in the spatial and channel dimensions by an adaptive fusion module.In addition,the model also embeds the doctor’s prior knowledge of disease observation tendency into the model learning process through an auxiliary attention guidance module.Finally,this thesis uses a heat map visualization technique to explain the prediction results of the model to help physicians better understand the process of neural network decision making and judge the reasonableness of the model decision.4.A prototype system for computer-aided diagnosis of intracranial vascular occlusion is designed and implemented.This system is based on the two model algorithms proposed in this thesis,which are packaged and provided to doctors for auxiliary diagnosis in the form of interface operations.