Table 1.

Machine-learning studies on stroke imaging

Application	Setting	Imaging tool	Performance
Diagnosis
Automatic lesion segmentation (ischemic stroke) [17]	Subacute stroke (> 24 hours and < 2 weeks)	MRI	Inferior to human segmentation
Automatic lesion segmentation (ischemic stroke) [18]	Chronic stroke	MRI (T1-weighted)	Comparable to manual segmentation
Automatic lesion segmentation (ischemic stroke) [19]	Acute stroke	DWI	Comparable to manual segmentation
Determination of ASPECTS (e-ASPECTS) [20,21]	Acute stroke	CT	Non-inferior to human reading
Automatic diagnosis of MCA dot sign [22]	Acute stroke (< 24 hours)	CT	Sensitivity 97.5%
Estimation of CSF volume for infarct edema [23]	Acute stroke	CT	Better than conventional method
Automatic lesion segmentation (hemorrhagic stroke) [25]	Acute stroke	CT	Comparable to manual segmentation
Prognosis
Symptomatic ICH after thrombolysis [27]	Acute stroke	CT	Improved the prognostic prediction
Improvement of visual function in PCA infarcts [28]	Subacute stroke (within 7 days)	MRI	Improved the prognostic prediction
Long-term mortality of AVM [34]	After endovascular treatment	CT, MRI	Accuracy of 97.5% to predict outcome
Impairment in multiple behavioral domains [35]	Subacute stroke (within 2 weeks)	MRI, fMRI	Enabled the prognostic prediction
Motor impairment [36]	Chronic stroke (≥ 3 months)	MRI, fMRI	Enabled the prognostic prediction

MRI, magnetic resonance imaging; DWI, diffusion weighted imaging; ASPECTS, Alberta Stroke Program Early Computed Tomography Score; CT, computed tomography; MCA, middle cerebral artery; CSF, cerebrospinal fluid; ICH, intracerebral hemorrhage; PCA, posterior cerebral artery; AVM, arteriovenous malformation; fMRI, functional magnetic resonance imaging.