Artificial IntelligenceArtificial intelligence in stroke imaging: Current and future perspectives
Introduction
Artificial intelligence (AI) is a subfield of computer science aiming at emulating human cognitive processes [1]. However, AI is an umbrella term that encompasses a number of techniques of varying complexity that can be used to solve different problems. One notable subfield of AI that has gained significant interest lately is machine learning (ML), including advanced deep learning methods [2]. Machine learning aims to automatically identify complex and high-dimensional patterns in existing datasets, which can then be used to make predictions or classifications based on new unseen datasets. Machine learning can be further subdivided into supervised and unsupervised machine learning as well as reinforcement learning. While all three types of machine learning can be useful for specific medical applications, supervised machine learning is arguably the most relevant machine learning type in this domain, especially with respect to precision medicine efforts. Briefly described, supervised machine learning methods are trained using “ground truth” labeled data to identify the corresponding patterns in the feature data [[3], [4], [5], [6]]. Examples of ML models include, for example, support vector machines (SVMs), decision trees, random forests (RFs), k-nearest neighbors, Bayesian methods, and generalized linear models (GLMs) [2]. Many of these methods can be used to solve classification (categorial outcome) as well as regression (continuous outcome) problems.
Lately, deep learning (DL) methods have gained increasing interest due to the superior performance achieved in many applications, including medicine, compared to other more traditional machine learning methods. Deep learning utilizes a specific type of a machine learning architecture, the so-called artificial neural networks, which loosely mimic how the human brain functions. Artificial neural networks consist of an input layer, multiple hidden layers, and one output layer (Fig. 1). Each layer consists of many neurons that are typically connected to neurons in the next layer [2,7,8]. Artificial neural networks use a set of quantitative variables as the input to the first layer, which are used together with corresponding labels to learn (optimize) the weights between the neurons in the network. Practically, these weights connecting the neurons are optimized through backpropagation and comparison of the predictions to the ground truth labels [2]. After optimization, the trained networks can be used to make predictions and classifications based on new unseen data.
Artificial neural networks cannot only use quantitative data as inputs, but also solve classification and regression problems on images that are used directly as an input to the network by using the intensity information of the pixels or voxels as input features. In this case, the artificial neural network is typically referred to as a convolutional neural network (CNN) [2,9]. Briefly described, the images or parts of images are analyzed in convolutional layers, which follow the input layer with the image and essentially contain filters that are optimized in the same way as the weights of the neural network using backpropagation to extract relevant texture information from the images. The quantified texture data is then fed through the hidden layers to make predictions on the image- or voxel-level (Fig. 2). CNNs have been shown to outperform classical machine learning techniques and image analysis methods in many computer vision problems. Radiology, and especially neuroradiology, is no exception to this, including applications for improved image quality, image acquisition and reconstruction, diagnostic accuracy, and treatment decision making for better patient care [1].
One pathology particularly amenable to machine learning advancements is stroke. Stroke is a leading cause of long term disability, morbidity, and mortality and timely intervention and increased sensitivity to subtle early findings can lead to significant benefits for the patients with respect to the clinical outcome [4]. Computed tomography (CT) and magnetic resonance imaging (MRI), including perfusion imaging (CT perfusion or perfusion-weighted MRI), play a critical role in diagnosis, intervention, and timely management of stroke [7]. Lately, AI has shown promise improving stroke management efficiency by incorporating its functionality in neuroimaging paradigm and workflow. This overview focuses on the recent advances of supervised ML and DL techniques, with respect to workflow, image acquisition and processing, and diagnosis pertinent to adult stroke management, while also discussing both the pitfalls and future directions within this subfield. This overview focusses specifically on CT and MRI applications due to the routine use of these neuroimaging techniques for diagnosis and treatment decision making, although a few supervised ML and DL techniques have been proposed for other neuroimaging techniques such as positron emission tomography and ultrasound imaging that can be of interest for specific stroke analyses too [[10], [11], [12]].
Section snippets
Imaging triage and workflow
Stroke is an acute neurological event where timely management is imperative. Thus, one area that has garnered interest is the prioritization of image interpretation by neuroradiologists. This is crucial since it ultimately expedites CT screening of hemorrhage or early signs of stroke through worklist prioritization. For example, Titano et al. [13] utilized a 3D CNN, which flags head CTs with high probability for acute neurologic events, including stroke, with an alarm mechanism for alerting
Image optimization
Another area in which AI has shown incredible progress lately is image optimization and quality improvement. These processes, which include fast image acquisition and reconstruction, denoising, artifact reduction, and resolution improvement, have considerable overlap in some instances. However, they will be discussed separately in this section as they pertain to stroke imaging.
Image analysis
Given the vast amount of data, time-sensitive nature of stroke, and potential for immediate clinical impact, image analysis has become arguably the most promising application of AI for stroke imaging. Image analysis encompasses detection and segmentation, classification, and prediction of stroke outcome using AI methods. These topics will be covered separately in this section despite exhibiting significant overlap in some cases.
Pitfalls/disadvantages
Supervised AI techniques have shown great promise for the support of image-based diagnosis support in acute ischemic stroke patients, but it is also essential to acknowledge the shortcomings and challenges, which prevent generalizability and clinical implementation. In this section, we address the main issues from data-related, medicolegal, and ethical perspectives.
Future direction
In this AI overview, we have covered a number of promising applications of AI pertinent to ischemic stroke diagnosis and treatment. Nevertheless, there is room for improvement and exciting opportunities ahead within this field. Integration of these AI platforms into clinical practice and how they affect neuroradiologists' workflow are the most pressing questions in the current environment. Areas of potential future exploration include AI integration in the electronic health record (EHR),
References (91)
- et al.
An overview of deep learning in medical imaging focusing on MRI
Z Med Phys
(2019) Arterial spin label imaging of acute ischemic stroke and transient ischemic attack
Neuroimaging Clin N Am
(2011)Artificial intelligence in stroke care: deep learning or superficial insight?
EBioMedicine
(2018)- et al.
aLow-dose CT via convolutional neural network
Biomed Opt Express
(2017) - et al.
Fusing multi-scale information in convolution network for MR image super-resolution reconstruction
Biomed Eng Online
(2018) - et al.
Ultra-high resolution in-vivo 7.0T structural imaging of the human hippocampus reveals the endfolial pathway
Neuroimage
(2015) - et al.
Fully automatic acute ischemic lesion segmentation in DWI using convolutional neural networks
NeuroImage Clin
(2017) - et al.
Quantitative prediction of acute ischemic tissue fate using support vector machine
Brain Res
(2011) - et al.
ISLES 2015-a public evaluation benchmark for ischemic stroke lesion segmentation from multispectral MRI
Med Image Anal
(2017) - et al.
Lesion segmentation from multimodal MRI using random forest following ischemic stroke
Neuroimage
(2014)
Automating ischemic stroke subtype classification using machine learning and natural language processing
J Stroke Cerebrovasc Dis
Applying instance-based techniques to prediction of final outcome in acute stroke
Artif Intell Med
Development and validation of a penumbra-based predictive model for thrombolysis outcome in acute ischemic stroke patients
EBioMedicine
Building machine learning models without sharing patient data: a simulation-based analysis of distributed learning by ensembling
J Biomed Inform
Applications of deep learning to neuro-imaging techniques
Front Neurol
Deep learning in neuroradiology
Am J Neuroradiol
Machine learning in acute ischemic stroke neuroimaging
Front Neurol
Deep learning guided stroke management: a review of clinical applications
J Neurointerv Surg
Deep into the brain: artificial intelligence in stroke imaging
J Stroke
A comparison of shallow and deep learning methods for predicting cognitive performance of stroke patients from MRI lesion images
Front Neuroinform
Artificial intelligence in radiology
Nat Rev Cancer
ISLES 2016 and 2017-benchmarking ischemic stroke lesion outcome prediction based on multispectral MRI
Front Neurol
Deep learning fully convolution network for lumen characterization in diabetic patients using carotid ultrasound: a tool for stroke risk
Med Biol Eng Comput
Ultrasound-based carotid stenosis measurement and risk stratification in diabetic cohort: a deep learning paradigm
Cardiovasc Diagn Ther
Classification of carotid artery intima media thickness ultrasound images with deep learning
J Med Syst
Automated deep-neural-network surveillance of cranial images for acute neurologic events
Nat Med
Automated critical test findings identification and online notification system using artificial intelligence in imaging
Radiology
Deep convolutional neural network for inverse problems in imaging
IEEE Trans Image Process
Image reconstruction by domain-transform manifold learning
Nature
Deep learning with domain adaptation for accelerated projection-reconstruction MR
Magn Reson Med
A machine learning approach to perfusion imaging with dynamic susceptibility contrast MR
Front Neurol
Deep learning enables reduced gadolinium dose for contrast-enhanced brain MRI
J Magn Reson Imaging
Recommended implementation of arterial spin-labeled perfusion MRI for clinical applications: a consensus of the ISMRM perfusion study group and the European consortium for ASL in dementia
Magn Reson Med
Improving arterial spin labeling by using deep learning
Radiology
Improvement of image quality at CT and MRI using deep learning
Jpn J Radiol
Stacked competitive networks for noise reduction in low-dose CT
PLoS One
A deep convolutional neural network using directional wavelets for low-dose X-ray CT reconstruction
Med Phys
Deep learning-based image quality improvement for low-dose computed tomography simulation in radiation therapy
J Med Imaging (Bellingham, Wash)
Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network
Jpn J Radiol
Deep learning based noise reduction for brain MR imaging: tests on phantoms and healthy volunteers
Magn Reson Med Sci
Convolutional neural network based metal artifact reduction in X-ray computed tomography
IEEE Trans Med Imaging
MoCoNet: Motion correction in 3D MPRAGE images using a Convolutional Neural Network approach. ArXiv: 180710831
Retrospective correction of motion artifact affected structural MRI images using deep learning of simulated motion
Med Imaging with Deep Learn
Computed tomography super-resolution using deep convolutional neural network
Phys Med Biol
Brain MRI super resolution using 3D deep densely connected neural networks
Cited by (32)
Evaluation of CINA® LVO artificial intelligence software for detection of large vessel occlusion in brain CT angiography
2024, European Journal of Radiology OpenProviding clinical context to the spatio-temporal analysis of 4D CT perfusion to predict acute ischemic stroke lesion outcomes
2024, Journal of Biomedical InformaticsPredicting treatment-specific lesion outcomes in acute ischemic stroke from 4D CT perfusion imaging using spatio-temporal convolutional neural networks
2022, Medical Image AnalysisCitation Excerpt :Nonetheless, using simple thresholds to identify the tissue status fails to capture the heterogeneous complexity of the AIS (Rekik et al., 2012; Flottmann et al., 2017) and does not provide any estimation of a treatment-specific outcome. Due to the success of machine learning, new approaches have been proposed to predict stroke lesion outcomes from medical image data (Yedavalli et al., 2021; Soun et al., 2021). Technically, stroke lesion outcome prediction consists of automatically estimating follow-up changes in the anatomical extent of lesions over time, based on imaging data acquired at admission to the hospital.
Brain stroke classification and segmentation using encoder-decoder based deep convolutional neural networks
2022, Computers in Biology and MedicineCitation Excerpt :One of the main purposes of artificial intelligence studies is to protect, monitor and improve the physical and psychological health of people [1].
Leveraging artificial intelligence in ischemic stroke imaging
2022, Journal of NeuroradiologyCitation Excerpt :Some of them are commercially available [e.g., RAPID ASPECTS (iSchemaView, Menlo Park, CA, USA): https://www.rapidai.com/].36 In a side-by-side comparison between the RAPID ASPECT score and the ASPECT score calculated by the neuroradiologists, the RAPID has better agreement with the final infarction on the DWI sequence.33,38–39 Frontier ASPECT Score Prototype (Siemens Healthcare GmbH, Erlangen, Germany) is another example of ASPECTS automation.