Artificial intelligence in neuroimaging with a focus on acute and degenerative neurologic disorders: a narrative review
Article information
Abstract
Purpose
Recent advancements in artificial intelligence (AI), especially in deep learning algorithms, have driven significant innovations across numerous industries, including medicine. Neuroimaging, faced with challenges from frequent acute neurological conditions and a rising prevalence of neurodegenerative disorders, has become an active field where AI is increasingly integrated into clinical workflows.
Current Concepts
In acute neurological disorders, AI models have been developed to improve the diagnostic accuracy of computed tomography and magnetic resonance imaging in detecting acute intracerebral hemorrhage and ischemic stroke. These systems expedite lesion identification, assist in patient triaging, and predict critical outcomes such as hematoma expansion from imaging features. Similarly, in neurodegenerative diseases such as Alzheimer dementia and Parkinson disease, AI enhances quantitative assessment of brain atrophy and identifies subtle imaging alterations that are challenging to detect visually. These AI solutions are now commercially available and already integrated into clinical practice. Surveys among neuroradiologists indicate growing acceptance of AI, acknowledging its potential to decrease workload and enhance clinical decision-making.
Discussion and Conclusion
Despite these promising advancements, clinical adoption faces challenges due to the need for standardized imaging protocols and AI systems capable of revealing new insights from conventional studies. Future efforts should focus on integrating AI into existing diagnostic workflows to provide innovative diagnostic insights, paving the way for personalized and effective patient care.
Introduction
1. Background
Recent developments in artificial intelligence (AI), particularly deep learning algorithms, have propelled innovations across various industries. Within healthcare, AI technologies have primarily been adopted in medical imaging specialties, including radiology, pathology, and ophthalmology. This has facilitated the creation and clinical deployment of AI-based products that support early disease diagnosis and treatment decision-making [1,2]. Additionally, recently developed large language models, such as OpenAI’s ChatGPT, have demonstrated considerable potential in automating electronic medical records management [3].
Neuroradiology frequently addresses neurological emergencies, necessitating high diagnostic accuracy due to the significant risks associated with interpretation errors. Furthermore, with aging societies, neurodegenerative diseases characterized by imaging findings such as brain atrophy are becoming more prevalent. These subtle degenerative changes can be challenging to quantify visually. Concurrently, the increasing demand for imaging studies, including computed tomography (CT) and magnetic resonance imaging (MRI), significantly burdens radiologists. Thus, active efforts have been made to integrate AI technologies to enhance diagnostic efficiency and reduce radiologists’ workloads [4,5]. Indeed, neuroimaging products represent the largest proportion of AI-based medical imaging devices authorized in Europe as of 2021 [2].
2. Objectives
This review examines current AI applications developed by Korean companies for acute and degenerative neurological disorders, and discusses existing challenges and future directions for AI in neuroradiology.
Applications of artificial intelligence in acute neurological disorders
1. Diagnosis of acute intracerebral hemorrhage
Acute intracerebral hemorrhage (ICH) is a severe condition requiring immediate diagnosis and intervention. Characteristic hyperdense signals observed on CT scans facilitate diagnosis, prompting the development of AI-based diagnostic tools for automatic hemorrhage detection [6,7]. Most commercial products include patient triaging systems, automatically alerting clinicians to potential hemorrhages, thus expediting diagnosis and treatment [8]. A recent study demonstrated that an AI model trained on CT scans from over 3,000 patients significantly improved diagnostic accuracy for acute hemorrhage compared to conventional interpretation (97.0% vs. 94.7%, P<0.0001) (Figure 1) [6].
Artificial intelligence (AI) software for detecting hemorrhage on brain computed tomography (CT). CT images of a 42-year-old male patient who presented with head trauma (A), along with intracranial hemorrhage detection results from AI software (B). Small amounts of acute subdural hemorrhage along the falx cerebri and subarachnoid hemorrhage along the left frontal sulcus are evident. The hemorrhage was accurately predicted by the AI software with a 98.4% probability.
One primary cause of ICH is aneurysm rupture. Non-invasive imaging techniques such as CT angiography (CTA) and MR angiography (MRA) allow early aneurysm detection, although small aneurysms remain challenging for radiologists to identify. To reduce this risk, various AI-assisted aneurysm detection tools have been introduced (Figure 2), demonstrating improved sensitivity among neurologists and neurosurgeons [9].
Artificial intelligence (AI) software for detecting cerebral aneurysms based on magnetic resonance angiography. Maximum intensity projection image (A) and source image (B) from time-of-flight magnetic resonance angiography. A cerebral aneurysm is seen in the cavernous segment of the right internal carotid artery, successfully detected by the AI model (indicated by a red circle and rectangle).
Predicting hematoma expansion is essential for determining appropriate treatment strategies. Radiologic markers, such as the “spot sign” on CTA or the “swirl sign” on non-contrast CT, indicate the risk of hematoma growth [10]. Recent studies reported enhanced predictive accuracy using AI models [11]. Although commercial solutions for predicting hematoma expansion are currently unavailable, future advanced AI developments may facilitate personalized therapeutic decisions based on prognosis.
2. Diagnosis of acute ischemic stroke
Time elapsed since symptom onset is crucial in managing acute ischemic stroke. Intravenous thrombolysis is typically administered within 4.5 hours, and endovascular treatment is considered within 6 hours when large vessel occlusion is confirmed [12]. Emerging evidence supports extending eligibility for endovascular treatment to patients presenting within 24 hours if ischemic penumbra is evident on perfusion imaging.
Non-contrast CT primarily excludes hemorrhage but can also identify early ischemic changes. CTA and MRA confirm large vessel occlusions, with CTA generally preferred for its rapid acquisition. MRI diffusion-weighted imaging sensitively detects even small ischemic lesions. Perfusion imaging, which differentiates ischemic core from penumbra regions, guides treatment decisions; however, its analysis involves complex calculations effectively managed by AI-based systems (Figure 3A–C) [13–16].
Diagnosis of acute ischemic stroke using artificial intelligence (AI) software. A case of a female patient in her 70s, illustrating AI software application across various imaging modalities for acute ischemic stroke diagnosis (A–C). (A) Non-contrast head computed tomography (CT) shows no evidence of acute intracerebral hemorrhage; however, AI software suggests hypodensity changes due to ischemic stroke. (B) CT angiography indicates occlusion of the left middle cerebral artery, with AI software reporting a large vessel occlusion score of 100. The occlusion was subsequently confirmed by cerebral angiography. (C) CT perfusion imaging shows ischemic core and penumbra volumes of 8.6 mL and 157.6 mL, respectively. Diffusion-weighted imaging acquired after endovascular treatment confirmed ischemic core volume expansion to approximately 50 mL. (D) Screenshots of a mobile application facilitating real-time AI analysis sharing and communication between paramedics and physicians. The interface includes clinical information, AI-generated results, and built-in secure chat functionality.
Recent Korean studies emphasize the importance of initial hospital selection in influencing treatment outcomes [17]. Patients transported directly to thrombectomy-capable hospitals experienced higher endovascular treatment rates (33.3%) compared to those admitted to primary stroke hospitals (12.2%) and showed significantly improved outcomes. AI-based mobile applications designed for rapid patient triaging and inter-provider communication (Figure 3D) demonstrate the potential to minimize transfer delays and improve clinical outcomes.
Stroke etiology, including cardioembolism, large artery atherosclerosis, and small vessel occlusion, significantly influences treatment strategies [18]. Determining stroke location through imaging is critical, and AI models have shown high concordance with expert evaluations [19]. AI also assists in differentiating large artery atherosclerosis from small vessel occlusion by accurately identifying stenotic lesions in relevant arteries [20]. AI applications estimating stroke onset time through fluid-attenuated inversion recovery (FLAIR) imaging signal intensity are currently under investigation [21].
Applications of artificial intelligence in Neurodegenerative disorders
1. Diagnosis and monitoring of Alzheimer disease
Brain atrophy is a hallmark imaging finding in dementia, varying according to subtype. In Alzheimer disease, atrophy prominently affects the medial temporal lobe, including the hippocampus. Although visual assessments of ventricular and sulcal enlargement are common practice, precise quantitative evaluations require dedicated volumetric software, particularly beneficial for follow-up imaging comparisons. Traditional volumetric software such as FreeSurfer previously required up to 10 hours per MRI scan. In contrast, recent deep learning-based solutions complete the analyses within seconds, substantially improving clinical efficiency [22,23].
The introduction of anti-amyloid therapies has necessitated precise detection of amyloid-related imaging abnormalities (ARIA). These abnormalities include the microhemorrhage- or superficial siderosis-related subtype (ARIA-H) and the edema- or effusion-related subtype (ARIA-E). AI can rapidly and accurately quantify these lesions, significantly reducing manual measurement errors and workloads (Figure 4) [24,25].
Artificial intelligence (AI)-based quantitative assessment of brain atrophy and detection of amyloid-related imaging abnormalities (ARIA). (A) Microbleeds detected in the initial study are marked with red boxes; new microbleeds found on follow-up imaging are highlighted with yellow boxes, facilitating rapid assessment of ARIA-H. (B) Initial fluid-attenuated inversion recovery (FLAIR) images display hyperintense white matter lesions in yellow; areas with increased lesion size on follow-up imaging are highlighted in red, aiding the evaluation of ARIA-E.
2. Diagnosis of Parkinson disease
Parkinson disease, characterized by the loss of dopaminergic neurons in the substantia nigra pars compacta, is typically diagnosed using nuclear medicine scans such as 123I-FP-CIT single-photon emission computed tomography (SPECT). Additionally, susceptibility-weighted MRI, which reveals loss of the nigrosome, aids in early diagnosis. AI models have significantly improved the detection rate of nigrosome loss, overcoming traditional imaging limitations and enhancing clinical applicability (Figure 5) [26,27]. While MRI cannot fully replace SPECT or positron emission tomography, further technological advances are expected to reduce radiation exposure and streamline clinical processes.
Artificial intelligence (AI)-based detection of nigrosome loss in the substantia nigra. (A) Image analysis results from a normal subject and (B) from a patient with Parkinson disease. In the susceptibility-weighted images from the Parkinson disease patient, the AI software correctly identifies nigrosome loss in the bilateral substantia nigra.
3. Diagnosis and monitoring of cerebral small vessel diseases
On MRI, cerebral small vessel diseases manifest as white matter hyperintensities, microbleeds, lacunes, and enlarged perivascular spaces. AI models facilitate accurate quantification and longitudinal monitoring of these lesions, which is crucial given their subtlety in visual inspection alone [28]. Numerous AI products support the detection and monitoring of these conditions, benefiting not only small vessel diseases but also conditions with similar imaging findings, such as multiple sclerosis.
Impact, challenges, and future directions
AI has substantially impacted neuroradiology by improving diagnostic prioritization, accuracy, and workload management for acute and degenerative neurological conditions [2,4]. Initial concerns about AI potentially replacing radiologists [29] have shifted towards widespread optimism. Surveys among Korean neuroradiologists highlight expectations that AI will enhance clinical workflow efficiency and diagnostic accuracy [5].
However, variability in imaging protocols across different devices complicates consistent AI performance, underscoring the need for standardized protocols and multicenter collaborations. Recent industry-driven advancements have begun to address these challenges, producing more robust AI solutions. Nonetheless, rigorous clinical validation remains essential to prevent diagnostic errors.
Current AI products primarily focus on efficiency and accuracy improvements rather than novel diagnostic insights. Future innovations should aim to generate clinically relevant new information, accompanied by cost-effectiveness analyses for insurance coverage and national-level policy support to improve global competitiveness.
Conclusion
Rapid diagnosis and prompt treatment are critical in acute neurological disorders such as ICH and ischemic stroke. AI applications enable near-real-time patient triaging, significantly enhancing emergency care. In degenerative neurological conditions, AI accurately detects subtle imaging changes and assesses therapy-related complications, enhancing diagnostic precision and clinical efficiency. Future advances offering novel diagnostic insights will further broaden AI’s clinical utility.
Notes
Conflict of Interest
Leonard Sunwoo is currently employed part-time at JLK Inc., Seoul, Korea. Otherwise, there are no conflicts of interest relevant to this article to declare.
Funding
None.