Neuroimaging and deep learning for brain stroke detection - A review of recent advancements and future prospects
Introduction
Stroke is one of the major reasons for adult deaths around the globe, impacting 6.2 million people every year [1]. Over the past two decades, there has been a 26 percent increase in stroke deaths, worldwide. Stroke is the second leading cause of death across the globe [2]. In turn, a great amount of research has been carried out to facilitate better and accurate stroke detection. Brain stroke occurs when the blood flow to the brain is stopped or when the brain doesn't get a sufficient amount of blood. As a result, the particular part of the brain drained of blood supply experiences a shortage of oxygen and becomes unresponsive [3]. This in turn, causes disturbances to the organs that are controlled by the affected part of the brain. The primary symptom of a stroke is partial numbness and numbness in the legs, arms and face belonging to one side of the body. Other minor symptoms are dizziness, headache, difficulty in walking and unconsciousness [4].
Depending on the obstacle in the blood supply to the brain, stroke can be classified into two types, Ischemic Stroke and Hemorrhagic stroke [5]. Ischemic stroke is caused by an obstruction in the blood vessels that carry blood to the brain. This specific type accounts for almost 87% of all stroke cases [6]. Hemorrhagic stroke is majorly caused by the breakage of weak blood vessels. Aneurysms and arteriovenous malformations are the basic types of weakened blood vessels that are responsible for Hemorrhagic stroke [7]. Another cause of Hemorrhagic stroke is high blood pressure [8].
The impact of stroke in an individual is influenced by the affected region of the brain and its severity. An extreme case could potentially lead to death [9]. The challenging process concerning stroke is its treatment. If the blood flow to the brain is not restored within several hours after the onset of ischemia, then the penumbral tissue will not be salvageable. The goal of stroke treatment and therapy is to save the penumbral tissue. Once the tissue is infarcted, the process is irreversible [10].
A stroke can be caused by a ruptured blood vessel or a blocked artery [11]. Images of the brain that are recorded during a scan and physical tests are utilized in diagnosing stroke among individuals. Strokes are diagnosed using advanced imaging techniques. Diagnosis is done with the help of brain imaging procedures such as Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) [12]. These imaging techniques have been proven to be essential for computing the changes in the properties of the tissues. The degree of infarction of the brain tissue is characterized by these imaging techniques. Over the past few years, various medical image investigation methodologies and statistical tools have been experimented with. The performance of these tools and methodologies in differentiating amongst the states of tissues in the brain were compared.
CT scans performed on an individual presumed to have a stroke, can be used to identify the type of stroke that they have undergone. The required information for carrying out the emergency procedures right after a stroke can be provided by CT scans [13]. CT is comparatively inexpensive and less affected by noise. CT is more accessible to patients and much faster than other imaging techniques such as MRI [14], [15]. In the event of a stroke, a non-enhanced CT is the first radiological examination performed on the patient [16]. A hypodense structure in the CT images indicates the presence of an ischemic lesion [17]. However, abnormal lesions are not clearly visible in a CT. It also experiences constraints while locating minute infarcts in the cerebellum, the brain stem and the interiors of the cerebral hemispheres.
Hence, MRI is a more suitable technique for overcoming this limitation [18]. Despite the many benefits that MRI provides, MRI is expensive and accessible to only a minority of the healthcare centers. The high duration for performing an MRI scan poses serious challenges to the existing imaging techniques [19]. By using MRI instead of CT, infarcts can be detected earlier on before the appearance of symptoms.
Over the last decade, several computer-aided techniques and tools have been developed to detect abnormalities in the brain at the earliest. Specifically, Artificial Intelligence and Deep Learning are majorly used to achieve accurate and automated results for Stroke detection. These results do not play a stand-alone role in the detection process. Being a very sensitive treatment process, computer-aided techniques can be complementary to support physicians in the stroke detection process [20].
There are certain challenges involved while using computer-aided techniques to get accurate results. Firstly, high-resolution images are required. The equipment required to get these high-resolution brain images is expensive. Secondly, the diversity in brain tissues increases the difficulty in diagnosing the stroke. Another challenge is the presence of any older stroke area. This makes the distinguishing process tough between the new stroke area and the old stroke area [21]. To get an efficient and accurate computer-aided model, all these issues need to be fixed.
Through this work, we aim to achieve the following objectives
- •
To compare the performance of the existing deep learning techniques used for stroke lesion detection and segmentation.
- •
To discuss the achievements of the state-of-the-art techniques in stroke lesion detection.
- •
To analyze the key challenges involved in stroke lesion detection and segmentation.
- •
To highlight the potential research gaps and future trends related to computer-aided diagnosis of brain stroke.
Section snippets
Search strategy and organization of the review
In this study, we have referred research publication databases like Pubmed™, ScienceDirect™, IEEEXplore™, and Google Scholar™ to search the relevant publications made in the area of brain stroke detection. In addition to these sources, few significant articles were also downloaded from Springer and Wiley publications. These articles were screened in the context of deep learning methods and the search strategy process is highlighted in Fig. 1.
We applied a three level filtering process to
Neuroimaging for stroke – a walkthrough
This section presents an overview of types of stroke and its related imaging modalities employed for diagnosis and treatment.
Deep learning for stroke detection
Though Machine learning methods were successfully applied in Medical image processing for the past two decades, it suffers from few limitations. These methods relied greatly on hand crafted features designed by domain experts. As the observed data vary from patient to patient and data interpretation varies with the experience of the domain experts, it might lead to intra and inter-observer error. On the other hand, deep learning in Medical imaging has made significant progress in capturing
Discussion
A comprehensive study has been carried out in this work by examining various deep learning approaches applied to different imaging modalities. To the best of our knowledge, this is the first report, presenting the modality-wise deep learning approaches employed exclusively for stroke lesion detection.
Conclusion
Deep learning models can never replace doctors and radiological experts. But it can make a huge impact on the automation process in image processing and analysis. Computer-aided techniques for analysis of medical images have grown significantly in recent times, contributing to medical research and clinical applications. Recent progress in deep learning has shown continuous optimization in the segmentation process of stroke lesion regions from the brain. This research aimed to observe the
Declaration of Competing Interest
None.
References (113)
- et al.
Use of nanoparticle contrast agents for cell tracking with computed tomography
Bioconjug. Chem.
(2017) - et al.
Brain ischemia: CT and MRI techniques in acute ischemic stroke
Eur. J. Radiol.
(2017) - et al.
Imaging of cerebral ischemia
Neurol. Clin.
(2014) - et al.
Imaging mass spectrometry revealed the production of lyso-phosphatidylcholine in the injured ischemic rat brain
Neuroscience
(2010) - et al.
Global changes in phospholipids identified by MALDI MS in rats with focal cerebral ischemia
J. Lipid Res.
(2012) - et al.
Fully automatic acute ischemic lesion segmentation in DWI using convolutional neural networks
NeuroImage
(2017) - et al.
Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation
Med. Image Anal.
(2017) - et al.
A deep supervised approach for ischemic lesion segmentation from multimodal MRI using fully convolutional network
Appl. Soft Comput.
(2019) - et al.
White matter hyperintensity and stroke lesion segmentation and differentiation using convolutional neural networks
NeuroImage
(2018) - et al.
Efficient multi-kernel DCNN with pixel dropout for stroke MRI segmentation
Neurocomputing
(2019)
Use of deep learning to predict final ischemic stroke lesions from initial magnetic resonance imaging
JAMA Netw. Open
Simulated perfusion MRI data to boost training of convolutional neural networks for lesion fate prediction in acute stroke
Comput. Biol. Med.
Automated segmentation of subarachnoid hemorrhages with convolutional neural networks
Inf. Med. Unlocked
A multi-path 2.5 dimensional convolutional neural network system for segmenting stroke lesions in brain MRI images
NeuroImage
Perioperative stroke
Can. J. Anaesth.
Incidence & prevalence of stroke in India: a systematic review
Indian J. Med. Res.
Ischemic Stroke, StatPearls [Internet]
Patient knowledge on stroke risk factors, symptoms and treatment options
Vascular Health and Risk Management
Stroke in the21stcentury: a snapshot of the burden, epidemiology, and quality of life
Stroke Res. Treat.
Hypertension and stroke: update on treatment
Eur. Cardiol. Rev.
Characteristics of hemorrhagic stroke following spine and joint surgeries
Biomed. Res. Int.
Therapeutic approach to hypertensive emergencies: hemorrhagic stroke
High Blood Press. Cardiovasc. Prevent.
Brain–heart interaction
Circ. Res.
Penumbral salvage and thrombolysis outcome: a drop of brain, a week of life
Brain
Simulation studies for non invasive classification of ischemic and hemorrhagic stroke using near infrared spectroscopy
Imaging of acute ischemic stroke
Emerg. Radiol.
Imaging of acute stroke prior to treatment: current practice and evolving techniques
Br. J. Radiol.
A “one-stop-shop” 4D CTA protocol using 320-row CT for advanced imaging in acute ischemic stroke: a technical note
Eur. Radiol.
Accuracy and reliability of the recommendation for IV thrombolysis in acute ischemic stroke based on interpretation of head CT on a smartphone or a laptop
Am. J. Roentgenol.
A survey on comparison analysis between EEG signal and MRI for brain stroke detection
Adv. Intell. Syst. Comput. Emerg. Technol. Data Mining Inf. Secur.
Computer-aided imaging analysis in acute ischemic stroke – background and clinical applications
Neurol. Res. Pract.
Ischemic stroke detection system with a computer-aided diagnostic ability using an unsupervised feature perception enhancement method
Int. J. Biomed. Imaging
Comparison of neurological clinical manifestation in patients with hemorrhagic and ischemic stroke
World J. Emerg. Med.
Guidelines for the primary prevention of Stroke
Stroke
Diagnosis and management of acute ischemic stroke: speed is critical
Can. Med. Assoc. J.
Acute ischemic stroke: current status and future directions
Mo. Med.
Effect of IV alteplase on the ischemic brain lesion at 24–48 hours after ischemic stroke
Neurology
Brain infarct segmentation and registration on MRI or CT for lesion-symptom mapping
J. Vis. Exp.
Imaging of ischemic stroke
CONTINUUM
Neuronal death after hemorrhagic stroke in vitro and in vivo shares features of ferroptosis and necroptosis
Stroke
Neuroimaging in Intracerebral Hemorrhage
Hemorrhagic Stroke
Magnetic resonance imaging of cerebral hemorrhagic stroke
Top. Magn. Reson. Imaging
Neuroimaging in intracerebral hemorrhage
Stroke
Distinguishing core from penumbra by lipid profiles using Mass Spectrometry Imaging in a transgenic mouse model of ischemic stroke
Sci. Rep.
Viability thresholds and the penumbra of focal ischemia
Ann. Neurol.
MALDI mass spectrometric imaging of lipids in rat brain injury models
J. Am. Soc. Mass Spectrom.
Visualization by mass spectrometry of 2‐dimensional changes in rat brain lipids, including N ‐acylphosphatidylethanolamines, during neonatal brain ischemia
FASEB J.
Imaging mass spectrometry detection of gangliosides species in the mouse brain following transient focal cerebral ischemia and long-term recovery
PLoS ONE
Accuracy of CT cerebral perfusion in predicting infarct in the emergency department: lesion characterization on CT perfusion based on commercially available software
Emerg. Radiol.
Cited by (59)
Classification of brain strokes using divergence-based convolutional neural networks
2024, Biomedical Signal Processing and ControlA comprehensive review and experimental comparison of deep learning methods for automated hemorrhage detection
2024, Engineering Applications of Artificial IntelligenceA few-shot learning-based ischemic stroke segmentation system using weighted MRI fusion
2023, Image and Vision ComputingDeep learning prediction of motor performance in stroke individuals using neuroimaging data
2023, Journal of Biomedical InformaticsAutomated ischemic acute infarction detection using pre-trained CNN models’ deep features
2023, Biomedical Signal Processing and ControlStrokeViT with AutoML for brain stroke classification
2023, Engineering Applications of Artificial IntelligenceCitation Excerpt :Kuo et al. (2019) proposed a patch-based Fully Convolutional Neural Network (PatchFCN), which takes into account the local features in the slices but again falls short of situations where the whole stack of CT Scans can be used as input. More details of various techniques applied for classification of brain stroke are covered in the Review Paper by Karthik et al. (2020) and Sirsat et al. (2020). For the sake of completeness, we have tabled a few significant studies in Table 1.