Progress in the application of artificial intelligence in acute ischemic stroke imaging
Acute ischemic stroke is an acute cerebrovascular disease caused by atherosclerosis and thrombosis, and is the main cause of disability in adults. How to objectively and accurately evaluate the condition of patients with ischemic stroke, such as the location of vascular occlusion, the area of infarction, the state of cerebral collateral circulation, etc., is very important for clinical diagnosis and treatment. A large amount of imaging data will be generated during the diagnosis of ischemic stroke, but due to the experience level and time limitations of physicians, it is often difficult to analyze these imaging data thoroughly, resulting in misdiagnosis and missed diagnosis.
Combining artificial intelligence (AI) technology with imaging technologies such as CT and MRI to build and train auxiliary diagnosis models is conducive to improving the accuracy of disease diagnosis and providing great help for clinical treatment decisions. At present, AI has been widely used in acute ischemic stroke imaging. It plays an important auxiliary role in infarct detection, Alberta stroke program early CT score (ASPECTS) grading, large vessel occlusion detection, image segmentation and patient prognosis prediction; at the same time, it also has the ability to self-correct and can continuously improve accuracy based on feedback.
This article gives an overview of AI technology commonly used in acute ischemic stroke imaging, and introduces the application of AI in CT plain scan, CT angiography (CTA), CT perfusion (CTP) and MRI of stroke.
1. AI Overview
As a branch of computer science, AI can simulate, extend and expand human thinking processes and intelligent behaviors. AI includes computer vision, natural language processing, machine learning (ML) and other technologies. Among them, ML is most commonly used in medical imaging, which relies on different algorithms to perform in-depth analysis of complex and diverse data. Traditional ML algorithms include support vector machine (SVM), linear regression, logistic regression, decision tree, random forest, Bayesian learning, etc., which are currently widely used in the field of scientific research. In the application process of ML, these traditional algorithms are more effective for simple tasks, but their performance is poor for some complex clinical problems.
Deep learning (DL) is a new branch of ML, and its performance is superior to traditional ML algorithms. DL can roughly imitate the function of the human brain using a specific type of ML architecture, namely artificial neural network, commonly known as convolutional neural network (CNN). CNN can input not only quantitative data, but also pixel or voxel information to solve image classification and regression problems. Therefore, DL combines low-level features to form more abstract high-level attribute categories or features to discover the characteristic distribution of data, group the data, and ultimately achieve the purpose of diagnosis.
2. Application of AI in CT plain scan of ischemic stroke
CT plain scan is mainly used to exclude bleeding and other non-ischemic lesions, preliminarily determine whether there is fresh infarction, and determine the location and range of infarction. Specific manifestations of acute ischemic stroke include hyperdense vessel sign (HDVS) on CT plain scan, that is, the density of the middle cerebral artery on the affected side is higher than that on the normal side, indicating occlusion of the middle cerebral artery. Takahashi et al. used a model based on the SVM algorithm to detect HDVS on CT plain scan, and the sensitivity of the model reached 97.5%.
Shinohara et al. studied 46 cases with HDVS and 52 cases without HDVS, and compared the diagnostic performance of the ML model with that of neuroradiologists for HDVS. The results showed that the accuracy of the ML model was 81.6%, and the area under the receiver operating characteristic curve (AUC) was 0.869, while the initial accuracy of the neuroradiologists was 78.8% and the AUC was 0.882; however, after referring to the results of the ML model, the accuracy of the neuroradiologists in the second test increased to 84.7%, and the AUC reached 0.932.
It can be seen that AI is expected to assist physicians in improving the diagnostic efficiency of HDVS. ASPECTS is a scoring scale for evaluating early anterior circulation ischemic changes. The score divides the middle cerebral artery supply area into 10 areas, and 1 point is subtracted for each area involved in the range of reduced density. Kuang et al. used MR diffusion weighted imaging (DWI) as the gold standard and used ML to build a model to evaluate the grading of ASPECTS. The results showed that the specificity of the model was 91.8% and the sensitivity was 66.2%.
Nagel et al. developed a commercial software (e-ASPECTS) that can realize automatic ASPECTS grading. The software has a high consistency with neuroradiologists in the identification of low-density lesions on CT plain scans of patients with acute ischemic stroke, and the accuracy is similar; however, when acute ischemic stroke patients have other brain parenchymal lesions (such as demyelination of white matter, old cerebral infarction, etc.), the accuracy of e-ASPECTS is not as good as that of neuroradiologists.
Another study on the automated ASPECTS scoring commercial software (RapidASPECTS) found that, with ASPECTS on DWI as the gold standard, the consistency between the automated software's ASPECTS grading based on CT plain scan and the gold standard (κ=0.9) was higher than that of neuroradiologists (κ=0.56~0.57). At 1 hour after the onset of acute ischemic stroke, the consistency between the automated software's ASPECTS grading and the gold standard was good (κ=0.78), and at 4 hours after the onset, the consistency between the two was even better (κ=0.92). For patients with acute ischemic stroke, clarifying the volume of the infarction helps guide clinical treatment and predict patient prognosis. AI has gradually achieved segmentation and measurement of infarction volume on CT plain scan images, such as CNN technology has been applied to the segmentation of infarction volume.
Sales et al. obtained the infarct volume based on CNN segmentation on CT plain scan images of acute ischemic stroke, which was highly consistent with the infarct volume obtained by DWI as the gold standard, and the intraclass correlation coefficient (ICC) reached 0.88 (the best ICC is 1).
3. Application of AI in CTA/CTP of ischemic stroke
CTA can clearly display the internal carotid artery, vertebral artery, basilar artery, anterior, middle and posterior cerebral arteries, and can be used to identify the responsible vessels of acute ischemic stroke and evaluate its collateral circulation. The use of AI technology in CTA images makes it possible to automatically detect large vessel occlusion, measure the infarct core and evaluate the collateral circulation in patients with ischemic stroke. Rodrigues et al. used AI commercial software (Via.ai) to detect and analyze 610 stroke patients. The results showed that the diagnostic sensitivity, specificity and accuracy of the software for bilateral internal carotid artery or anterior and middle cerebral artery occlusion were 87.6%, 88.5% and 87.9%, respectively.
Sheth et al. used the symmetry of the brain to develop RAPID software to identify occlusion of bilateral internal carotid arteries or anterior and middle cerebral arteries on CTA and measure the core infarction volume. The RAPID software evaluated the core infarction volume of 179 patients and used the core infarction volume on CTP as the gold standard. The results showed that the accuracy of the software was similar to that of CTP and the AUC was 0.88. You et al. combined CT scans with basic clinical information to construct a diagnostic model based on ML methods (including logistic regression, random forest, SVM, etc.), which was then verified in 300 patients with bilateral internal carotid artery or anterior and middle cerebral artery occlusion, and obtained a high diagnostic efficacy (AUC of 0.847). Oman et al. used three-dimensional CNN technology on CTA to detect and segment the core infarction area, and obtained an AUC value of 0.93 and a Dice coefficient of 0.61.
In addition, in the automated assessment of collateral circulation in ischemic stroke, Grunwald et al. used the commercial software e-CTA to evaluate the collateral circulation TAN score of 98 patients undergoing mechanical thrombectomy. The results showed that e-CTA increased the ICC of neuroradiologists from 0.58 to 0.77, and the sensitivity and specificity of the software for collateral circulation were 0.99 and 0.94, respectively. CTP can display the core infarct area.
In a retrospective study based on CTP images within 6 hours of ischemic stroke onset, the core infarct volume on DWI was used as the gold standard. The results showed that the accuracy of delay and dispersion correction in CNN was higher than that of the standardized process. Kasasbeh et al. used CNN to build a model based on CTP and clinical data to predict the core infarct volume of patients with acute ischemic stroke. DWI was also used as the gold standard. The results showed that the AUC of the CTP group alone was 0.85, the AUC of CTP combined with clinical data was 0.87, and the maximum Dice coefficient was 0.48. These studies show that the combination of AI and CTP can achieve accurate assessment of core infarct volume.
4. Application of AI in MRI of ischemic stroke
For acute ischemic stroke, AI automatic segmentation and determination of the core infarct area on MRI images often use DWI images or apparent diffusion coefficient (ADC) maps, and the core infarct volume on DWI is often used as the gold standard. Kim et al. used a U-net model of encoding and decoding CNN to perform image segmentation on DWI and ADC maps, and found that the segmentation results of the U-net model were highly consistent with the manual segmentation results of experts, with an ICC as high as 1.0. Wu et al. used 3DCNN to segment acute ischemic lesions on DWI data from multiple centers, and compared the results with the manually measured lesion volumes, and found that the correlation between the two was excellent (correlation coefficient was 0.92). In addition to DWI sequences, MR perfusion imaging (perfusion weighted imaging, PWI) can also be used for stroke diagnosis and infarct segmentation.
Bouts et al. compared five DL algorithms in a rodent model with spontaneous or induced reperfusion and no reperfusion after right middle cerebral artery occlusion. Among them, the generalized linear model performed best in detecting the ischemic penumbra on PWI, with a Dice coefficient of 0.79. Huang et al. segmented the ischemic penumbra on the cerebral blood flow map and ADC map of PWI based on SVM to detect middle cerebral artery occlusion. The AUC of this model was 88% and 94% at 30 minutes and 60 minutes after arterial occlusion, respectively. In addition, some results have been obtained in the study of image segmentation of ischemic penumbra and core infarct area using multimodal MRI or fluid-attenuated inversion recovery (FLAIR) sequences. The multimodal MRI model combined with the random forest algorithm has a high accuracy in perfusion estimation and segmentation of infarct lesions, with average Dice coefficients of 0.82 and 0.59, respectively.
In the study of using FLAIR sequence for infarct segmentation, the accuracy of random forest model was also high, with Dice coefficient ranging from 0.54 to 0.6. The onset time of acute ischemic stroke is crucial for the choice of treatment, but sometimes it is impossible to accurately know the onset time of the patient. Lee et al. included some patients with acute ischemic stroke with unclear onset time, and used DWI-FLAIR mismatch to determine whether their onset time was within 4.5h, in order to determine whether intravenous thrombolysis was suitable. The results showed that the sensitivity of manual judgment of DWI-FLAIR mismatch was only 48.5%, while the judgment based on ML had higher sensitivity, such as the sensitivity of random forest was 72.7%, and the sensitivity of Logistic regression and SVM was 75.8%, and these three ML methods did not reduce the specificity. Therefore, AI can assist in determining the onset time of acute ischemic stroke patients and help the implementation of clinical decision-making.
Post-infarction hemorrhagic transformation is a serious complication after the treatment of ischemic stroke and is associated with poor prognosis. Yu et al. used 24-hour follow-up MRI as the gold standard and used ML models including SVM, linear regression, and decision trees to predict the site of potential hemorrhagic transformation on DWI and PWI images. The accuracy of the ML model was as high as 84%. Bouts et al. found that the random forest model based on multi-parameter MRI can effectively predict hemorrhage in an animal model of hemorrhagic transformation induced by reperfusion after infarction, with an AUC of 0.85-0.89. Ischemic stroke can lead to loss of cognitive or motor function. Therefore, the prediction of clinical prognosis of stroke patients can affect physicians' treatment decisions, help reduce additional complications of stroke, and maximize the quality of life of patients.
Recent studies have focused on using AI to predict the clinical prognosis of stroke patients in order to support early clinical decision-making. For example, Tang et al. developed and verified ML models based on algorithms such as SVM and decision trees to predict the short-term and long-term clinical prognosis of patients. They combined preoperative DWI and PWI imaging data and basic clinical data to build a comprehensive model, and compared it with a model with only basic clinical data and ischemic penumbra volume. It was found that the comprehensive model was more accurate in predicting short-term and long-term clinical prognosis results, with an AUC of 0.863.
A retrospective study used a gradient boosting algorithm to predict patient prognosis, combining imaging data, epidemiological data and basic clinical data to build a comprehensive model, and its accuracy in predicting poor prognosis was 87.7%. Chauhan et al. input 3DMRI imaging data into a CNN-based model to assess the severity of speech disorders in patients with ischemic stroke. By comparing ML models based on linear regression and SVM algorithms with CNN models, they found that the CNN model had the best ability to predict speech disorders, which was also verified in a larger sample size. Rehme et al. used functional MRI to discover neuroimaging markers of movement disorders related to stroke. They compared 40 stroke patients with 20 healthy subjects and constructed a model using the SVM algorithm on the DWI before treatment and the functional MRI after treatment. They found that the model can accurately identify stroke patients with and without hand movement disorders with an accuracy of up to 88%.
5. Summary
Combining AI with a variety of imaging examination methods can help improve the diagnostic efficiency of stroke. However, there are still some limitations in the use of AI in the study of acute ischemic stroke. First, AI cannot distinguish old lesions in the patient's brain parenchyma, and its accuracy needs to be further improved. Secondly, the current research on AI in ischemic stroke is mostly limited to the anterior circulation, and less is applicable to the posterior circulation. It is hoped that the software for detecting ischemic stroke can be further developed, improved and promoted in the future. Thirdly, there are still some challenges in the application of AI. Since the AI model needs to be trained with a large sample of data to obtain stable performance, how to build large-scale, high-quality training set data is a major difficulty of AI.
In addition, the accuracy of AI needs to be verified, which requires a lot of manpower and time. Finally, the legitimacy of data, especially the ethical protection measures in terms of data supervision, data privacy and network security, are the focus of the AI field. I believe that with the further combination of AI technology and medical imaging technology in the future, it will provide more powerful help for the diagnosis and treatment of ischemic stroke patients.