Professional Documents
Culture Documents
TOPICAL REVIEW
Artificial Intelligence Applications in Stroke
Kim Mouridsen , PhD; Patrick Thurner, MD; Greg Zaharchuk, MD, PhD
large vessel occlusion. One important caveat about stroke features, a deep neural network trained on a dataset with
imaging is that it must be done quickly, as faster treatment
known outputs can learn the best features for organiz-ing
leads to better outcomes.1 However, most steps in the
the data. In this review, we will discuss ML methods
stroke imaging triage pathway require the presence of
applied to stroke imaging with an emphasis on DL appli-
human radiologists and neurologists, and this is often the
cations. We refer to Figure for a graphical overview of the
time-limiting step. The expertise required for these tasks
applications discussed in this review.
may not be available at all sites or at all times. Therefore,
there is interest in automated methods for stroke imaging
evaluation.
FROM LINEAR THRESHOLDS TO
Artificial intelligence (AI) is a broad term reflecting the
use of computers to perform tasks that humans may find DEEP LEARNING
difficult, often in ways that are hard to pinpoint. For Early approaches to image analysis relied on apply-ing
example, although humans find high-level computa-tion fixed threshold values to quantitative maps to pre-dict
difficult, calculator technology is not considered AI information of interest, such as the amount of infarcted
because we know how to break this down into discrete tissue on follow-up or the location of hemor-rhagic
steps and feel we understand it. However, facial recogni- transformation. While simple to implement, apply-ing
tion is a task that humans perform well, but an algorithm to threshold values can fail due to patient cohort and scan
identify faces is usually considered AI since we cannot heterogeneity as well as post-processing variations.
articulate precisely how this is done. Machine learning More fundamentally, the underlying assumptions are
(ML) is a subset of AI in which algorithms learn from usually physiologically too simplistic as tissue outcome
Key Words: artificial intelligence deep learning hemorrhage ischemia machine learning
Correspondence to: Kim Mouridsen, PhD, CFIN, Aarhus University Hospital, Bldg 1A, 1st Floor, Nørrebrogade 44, DK-8000 Aarhus C, Denmark. Email kim@cfin.au.dk
For Sources of Funding and Disclosures, see page 2578.
© 2020 American Heart Association, Inc.
Stroke is available at www.ahajournals.org/journal/str
principles. A particular challenge with voxel-level have been raised about time delays associated with brain
prediction is that the characteristics of surround-ing tissue scanning. For this reason, arguments have been made
likely also have an influence. Differences in image-based against the use of advanced imaging such as CT perfu-
algorithm performance are due to both the imaging feature sion or MRI, which tend to take longer to acquire and
choices and the methods on how to combine information process than noncontrast CT and CT angiography. Most
about their spatial relationship. More recently, other ML MRI stroke protocols require <10 minutes of scan time,
methods have been proposed that go beyond the limits of although MRI does require more detailed screening and
regression, particularly those imposed by linear models. patient transfer time than CT. Thus, reducing scan time
These include support vector machines, classification with MRI continues to be important and is desirable given
trees, random forests, and dis-criminant analysis that can the high tissue contrast of DWI and its specificity for tissue
combine features in nonlinear highly likely to be irreversibly damaged.
AI-inspired methods can improve the quality of medi-cal linear combination of perfusion markers including
images, allowing them to be acquired faster or at lower cere-bral blood flow, blood volume, and Tmax.
Topical Review
dose. In particular, reconstruction algorithms based on
deep convolutional neural networks (CNNs) are well-
suited to this task. One example with relevance to acute PREDICTING CONTEMPORANEOUS
stroke imaging is the ability to use a CNN to de-noise MR INFORMATION
brain perfusion images using arterial spin labeling,
allowing diagnostic images to be created with shorter Time From Onset
scans.3 This approach has been applied to other MR The mismatch between DWI and FLAIR lesions has been
sequences as well, including quantitative susceptibility suggested as radiological indicator of whether a stroke is
mapping, which can detect brain hemorrhage and calci- more than 3 to 4 hours old. 10 If a lesion is visible on DWI but
fication.4 Given concerns regarding gadolinium deposi-tion not on FLAIR, the stroke is believed to have occurred more
following MR contrast administration, an AI method that recently. In practice, detecting the presence of a DWI-FLAIR
can significantly reduce the amount of gadolinium required mismatch visually can be challenging. Alter-natively, time from
for diagnostic imaging may find application for bolus onset can be directly estimated with machine learning as
perfusion-weighted imagin.5 While it is difficult to described in a study using a combina-tion of neural networks
significantly shorten acquisition times given the speed of and classical ML techniques.11 In a first step, features from
modern multidetector CT scanners, CNNs can be used to the MR source perfusion images were identified using a
reduce radiation dose.6 This may be helpful to allay neural network–based technique known as an auto-encoder.
radiation concerns related to CT perfusion imaging, a In the second step, these fea-tures were combined with DWI
relatively high-dose procedure.7 and FLAIR images using a different supervised learning
Another approach is to use DL to predict gold stan-dard algorithm (support vector machines and stepwise regression).
imaging biomarkers in situations where they are infeasible In a population of 105 patients with stroke, the authors report
to obtain. One example is that of cerebral blood flow an area-under-the-curve (AUC) of up to 0.68 to classify
imaging, where the reference standard is 15O-water patients into before/ after 4.5 hours from symptom onset.
positron emission tomography. Given the 2-minute half-life Another group used a similar approach including multimodal
of 15O, this requires an on-site cyclo-tron and 24-7 MRI for classifying patients into before/after 6 hours using
synthesis capability, which limits its use for acute diseases logistic regression showed equivalent performance.12
such as stroke. However, it is pos-sible to train a deep
Downloaded from http://ahajournals.org by on November 16, 2020
used for detection of intracranial large vessel occlusion. This compared to 0.43 and 0.79 for a standard
Results17 indicate a 82% sensitivity and 94% specificity single-network approach using a U-Net, 24 a popular
Topical Review
in a 650 person study. In practice, this has been shown choice for image segmentation.
to offer the possibility of early alerting of a senior physi- While MR manual delineation of core and at-risk tissue
cian and aid with task prioritization. can be used as target for AI algorithms, manual identifi-
cation of core lesions with CT perfusion is more difficult.
One way to train such a network is on paired sets of CT
Short-Term Tissue Segmentation: and MRI acquired in close temporal proximity. This was
Identification of Core/At-Risk Tissue the task for the Ischemic Stroke Lesion Segmenta-tion
Considerable work has been reported on automated (ISLES) 2018 Challenge, where the top-performing
identification of acute ischemic lesions (this section) and algorithm25 used a 3D neural network architecture to
lesion evolution (the following section). Two important achieve a Dice score of 0.56. Training (63 patients) and
types of tissue worth identifying on acute stroke stud-ies validation (40 patients) data was from patients present-ing
are (1) irreversibly damaged tissue (core lesion) and within 8 hours of symptom onset and with MR DWI
(2) the ischemic penumbra (at-risk tissue). As these acquired within 3 hours of the CT perfusion.
lesions have a physiological rather than imaging defini- There is a clear trend that lesion identification and
tion, establishing a reference against which to train and prediction of short-term tissue outcome is increasingly
validate automated methods is challenging. In practice, 2 often attempted with DL.26 This is clearly seen with the
approaches are used: manual delineation on acute imag- recent ISLES Challenges (http://www.isles-challenge.
ing or on follow-up imaging. The former approach aims for org/). These are open competitions where participants
an exact match between radiologist segmentation and train algorithms on standardized datasets. Performance
automated identification but assumes that tissue that is is evaluated on an independent test set where the out-
irreversibly damaged or at risk of infarction in the absence comes remain unknown to the participants. The com-
of reperfusion can be unambiguously outlined, which in petition has focused on lesion detection in MR (2015,
practice is challenging. Using follow-up imaging, the 2016), CT (2018), and the prediction of short-term tis-
assumption is that the core corresponds to the final tissue sue outcome (2019). Such competitions offer a natural
outcome in patients with successful recanaliza-tion and benchmark for algorithm performance and mitigates the
the at-risk lesion corresponds to the final tissue outcome problem of comparing performance between algorithms
in patients with persistent occlusion. Of course, most
Downloaded from http://ahajournals.org by on November 16, 2020
of 0.85±0.07 was demonstrated in 161 patients in a rtPA-treated cases (16 mL) was significantly lower than for
study based on CT perfusion and CT angiography untreated patients (29 mL). A more recent study eval-
Topical Review
imag-ing <7 hours after onset to predict tissue outcome uated a larger cohort of patients with LVOs. 35 As a first
after 2 to 7 days.29 However, tissue infarction likely step, the authors trained the network without knowledge of
depends on interactions between imaging markers; that reperfusion, yet still achieved better performance in both
is, the effect of a change in blood volume might be minimal and maximal reperfusion patients than by using
modu-lated by concurrent changes in flow. Linear current standard-of-care thresholding software. In addition,
models can handle such interactions, although this has this allowed prediction in patients with partial or unknown
been rarely exploited. With logistic regression, reperfusion, which comprise a large num-ber of patients
interactions must be explicitly stated before analysis, with stroke. While it is nonintuitive that DL can predict
which becomes inef-ficient with more complex tissue stroke outcomes regardless of perfusion status
interactions and as the num-ber of predictors increase. better than current methods that take this into account,
Tree-based methods offer a natural framework for there may be information on the ini-tial images that is
automatically identifying modulations of predic-tor related to the likelihood of subsequent reperfusion. One
effects.30 The computationally optimized XGBoost31 approach that maximizes the use of the available patient
extends previously discussed tree-based models. This data is one in which a model trained on partial or unknown
algorithm builds a large number of small classification reperfusion status can then be fine-tuned using smaller
trees with comparatively low predictive performance but number of cases of minimal and maximal reperfusion, a
subsequently combines the weak classifers into a stronger powerful method known as trans-fer learning (described in
one. In one report31 using data from the Euro-pean I-Know more detail below).
study32 and the ischemic per-conditioning study, 28 2 A fundamental limitation of data-hungry methods such
models were trained separately for recanaliz-ing and as DL is the inherent lack of large imaging datas-ets. This
nonrecanalizing patients based on perfusion-weighted is often mitigated by the use of either pretrain-ing or data
imaging and DWI. Predictive performance in the 2 groups augmentation. With pretraining, a deep neural network is
was good, with AUC 0.89 and 0.90, respec-tively, in the first trained on a larger dataset of typically nonmedical
recanalized and nonrecanalized patients. images. Then, using the settings obtained from this
Neural networks further increase the flexibility in pre- training procedure as initial values, the network is adapted
dicting outcome by simultaneously allowing nonlinearity to the particular medical setting, an approach known as
and automatically constructing predictive features
Downloaded from http://ahajournals.org by on November 16, 2020
stroke centers who experienced hemorrhagic transfor- studies have used retrospective data with sample sizes
mation after reperfusion therapy.38 Sensitivity and speci- ranging typically from 20 to a few hundred. There is a
Topical Review
ficity for hemorrhagic transformation location prediction in clear need for larger and especially prospective evalua-
anterior circulation was 89% and 60%, respectively, with tions building on the successful proof-of-concept
an AUC of 0.88. For CT, most studies used classical ML reports. While much work has yet to be done, some of
methods, such as the support vector machine. For the meth-ods reviewed here have obtained regulatory
instance, in one study39 with 116 rtPA patients, an AUC of clearance and are commercially available.
0.62 was demonstrated. This performance was com- While most attention currently is devoted to increas-
parable to the radiological SEDAN score (AUC 0.63), but ing predictive accuracy, endowing AI solutions with the
lower than the HAT score (AUC 0.72). Interestingly, with ability to account for their predictions would likely facili-
the addition of the clinical feature of the acute NIHSS, ML tate even better clinical applicability and acceptance.
performance increased significantly (AUC 0.74). Leveraging AI to identify new disease mechanisms and
unrealized connections between imaging and clini-cal
findings predictive of short- and long-term outcome
PREDICTION OF LONG-TERM OUTCOMES would allow AI to accelerate patient management and
While considerable attention has been afforded to lesion increase safety. It may also serve as means of
identification and short-term imaging progression, long-
hypothe-sis generation—paving the way to real deep
learning and understanding of acute ischemic stroke.
term functional outcome is arguably most significant.
Random forests have been used to predict functional
independence, defined as 90-day modified Rankin Scale ARTICLE INFORMATION
score ≤2, and good reperfusion, defined as modified TICI
Affiliations
≥2b.40 In 1383 patients from the MR CLEAN Registry, the Center of Functionally Integrative Neuroscience and MINDLab, Institute of
authors demonstrated poor performance to pre-dict Clini-cal Medicine, Aarhus University, Denmark (K.M.). Universitätsklinik für
reperfusion (AUC, 0.53–0.57). However, functional Radiologie und Nuklearmedizin, Vienna General Hospital, Austria (P.T.).
Diagnostic Radiology, Stanford University School of Medicine, CA (G.Z.).
independence could be moderately well predicted using
baseline variables only (AUC, 0.77–0.79) but well pre- Disclosures
dicted when treatment variables were included (0.88– Dr Mouridsen participates in Cercare Medical, equity, and salary concerns. Dr
Thurner participated in Cercare Medical and consultancy. Dr Zaharchuk received
0.91). While random forests performed optimally, other ML funding support through National Institutes of Health, GE Healthcare, and Bayer
Downloaded from http://ahajournals.org by on November 16, 2020
techniques such as logistic regression yielded vir-tually Healthcare, received GPU donation from Nvidia, and equity from Subtle Medical.
identical performance. A random forest approach using
the computationally optimized XGBoost version has also
demonstrated good performance41 in predict-ing good 90- REFERENCES
day clinical outcome. In 512 patients who underwent CT, 1. Saver JL. Time is brain–quantified. Stroke. 2006;37:263–266. doi:
10.1161/01.STR.0000196957.55928.ab
CTA, and CT perfusion imaging, an AUC of 0.75 for 2. Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with
prediciting day 90 modified Rankin Scale score ≤2 was deep convolutional neural networks. Proc NIPS. 2012; 1097–1105.
demonstrated. The inclusion of NIHSS at 24 hours 3. Kim KH, Choi SH, Park SH. Improving arterial spin labeling by using deep
learning. Radiology. 2018;287:658–666. doi: 10.1148/radiol.2017171154
increased performance to an AUC of 0.89. 4. Yoon J, Gong E, Chatnuntawech I, Bilgic B, Lee J, Jung W, Ko J,
In 204 patients, a combination of neural networks was Jung H, Setsompop K, Zaharchuk G, et al. Quantitative susceptibility
trained to predict improvement in the NIHSS of ≥4 points at 24 mapping using deep neural network: QSMnet. Neuroimage.
2018;179:199–206. doi: 10.1016/j.neuroimage.2018.06.030
hours and a modified Rankin Scale score of 0 to 1 at 90 5. Gong E, Pauly JM, Wintermark M, Zaharchuk G. Deep learning
days.42 A CNN was trained on noncontrast CT data, and a enables reduced gadolinium dose for contrast-enhanced brain MRI. J
Magn Reson Imaging. 2018;48:330–340. doi: 10.1002/jmri.25970
fully connected feed-forward neural network with 2 hidden
6. Chen H, Zhang Y, Zhang W, Liao P, Li K, Zhou J, Wang G. Low-dose
layers was trained with clinical data including blood pressure CT via deep neural network. Biomed Opt Express. 2017;8:679–694.
and NIHSS at presentation. In combination, they provided doi: 10.1364/ BOE.8.000679
7. Xiao Y, Liu P, Liang Y, Stolte S, Sanelli P, Gupta A, Ivanidze J, Fang
accuracies of 71% (for NIHSS improvement) and 74% (for
R. STIR-Net: Deep Spatial-Temporal Image Restoration Net for
good clinical outcome). For comparison, 3 clinical prognos-tic Radiation Reduction in CT Perfusion. Front Neurol. 2019;10:647. doi:
scores (HIAT, THRIVE, and SPAN-100) achieved maxi-mum 10.3389/ fneur.2019.00647
8. Guo J, Gong E, Fan AP, Goubran M, Khalighi MM, Zaharchuk G. Predicting
accuracies of 65% (NIHSS improvement) and 68% (good (15)o-water PET cerebral blood flow maps from multi-contrast MRI using a
clinical outcome), emphasizing that ML methods can deep convolutional neural network with evaluation of training cohort bias.
outperform existing outcome prediction tools. JCBFM. 2019. doi: 10.1177/0271678X19888123
9. Ho KC, Scalzo F, Sarma KV, Speier W, El-Saden S, Arnold C. Predicting
ischemic stroke tissue fate using a deep convolutional neural network on
source magnetic resonance perfusion images. J Med Imaging (Bellingham).
FUTURE CONSIDERATIONS 2019;6:026001. doi: 10.1117/1.JMI.6.2.026001
10. Thomalla G, Cheng B, Ebinger M, Hao Q, Tourdias T, Wu O, Kim JS,
There are many clinically valuable opportunities for AI in Breuer L, Singer OC, Warach S, et al; STIR and VISTA Imaging Inves-
the acute and subacute phases of ischemic stroke. Most tigators. DWI-FLAIR mismatch for the identification of patients with
acute ischaemic stroke within 4·5 h of symptom onset (PRE-FLAIR): a on multispectral MRI. Front Neurol. 2018;9:679. doi: 10.3389/fneur.
multicentre observational study. Lancet Neurol. 2011;10:978–986. doi: 2018.00679
10.1016/S1474-4422(11)70192-2 27. Wu O, Koroshetz WJ, Ostergaard L, Buonanno FS, Copen WA,
Topical Review
11. Ho KC, Speier W, El-Saden S, Arnold CW. Classifying acute ischemic Gonzalez RG, Rordorf G, Rosen BR, Schwamm LH, Weisskoff RM, et
stroke onset time using deep imaging features. AMIA Annu Symp al. Predicting tissue outcome in acute human cerebral ischemia using
Proc. 2017;2017:892–901. combined diffu-sion- and perfusion-weighted MR imaging. Stroke.
12. Wouters A, Dupont P, Christensen S, Norrving B, Laage R, Thomalla G, Kemp S, 2001;32:933–942. doi: 10.1161/01.str.32.4.933
Lansberg M, Thijs V, Albers GW, et al. Multimodal magnetic resonance imaging to 28. Hougaard KD, Hjort N, Zeidler D, Sørensen L, Nørgaard A, Hansen TM,
identify stroke onset within 6 h in patients with large vessel occlu-sions. Eur von Weitzel-Mudersbach P, Simonsen CZ, Damgaard D, Gottrup H, et al.
Stroke J. 2018;3:185–192. doi: 10.1177/2396987317753486 Remote ischemic perconditioning as an adjunct therapy to throm-bolysis in
13. Schröder J, Thomalla G. A critical review of alberta stroke program patients with acute ischemic stroke: a randomized trial. Stroke.
early CT score for evaluation of acute stroke imaging. Front Neurol. 2014;45:159–167. doi: 10.1161/STROKEAHA.113.001346
2016;7:245. doi: 10.3389/fneur.2016.00245 29. Kemmling A, Flottmann F, Forkert ND, Minnerup J, Heindel W,
14. Nagel S, Sinha D, Day D, Reith W, Chapot R, Papanagiotou P, Warburton Thomalla G, Eckert B, Knauth M, Psychogios M, Langner S, et al.
EA, Guyler P, Tysoe S, Fassbender K, et al. e-ASPECTS software is non- Multivariate dynamic prediction of ischemic infarction and tissue
inferior to neuroradiologists in applying the ASPECT score to computed salvage as a function of time and degree of recanalization. J Cereb
tomogra-phy scans of acute ischemic stroke patients. Int J Stroke. Blood Flow Metab. 2015;35:1397– 1405. doi: 10.1038/jcbfm.2015.144
2017;12:615– 622. doi: 10.1177/1747493016681020 30. Hastie T, Tibshirani R, Friedman JH. The elements of statistical
15. Guberina N, Dietrich U, Radbruch A, Goebel J, Deuschl C, Ringelstein A, learning: Data mining, inference, and prediction. Springer; 2009.
Köhrmann M, Kleinschnitz C, Forsting M, Mönninghoff C. Detection of early 31. Livne M, Boldsen JK, Mikkelsen IK, Fiebach JB, Sobesky J,
infarction signs with machine learning-based diagnosis by means of the Mouridsen K. Boosted tree model reforms multimodal magnetic
Alberta Stroke Program Early CT score (ASPECTS) in the clinical routine. resonance imag-ing infarct prediction in acute stroke. Stroke.
Neuroradiology. 2018;60:889–901. doi: 10.1007/s00234-018-2066-5 2018;49:912–918. doi: 10.1161/STROKEAHA.117.019440
16. Amukotuwa SA, Straka M, Smith H, Chandra RV, Dehkharghani S, Fischbein NJ, 32. Alawneh JA, Jones PS, Mikkelsen IK, Cho TH, Siemonsen S, Mouridsen K,
Bammer R. Automated detection of intracranial large vessel occlu-sions on Ribe L, Morris RS, Hjort N, Antoun N, et al. Infarction of ‘non-core-non-pen-
computed tomography angiography: a single center experience. Stroke. umbral’ tissue after stroke: multivariate modelling of clinical impact. Brain.
2019;50:2790–2798. doi: 10.1161/STROKEAHA.119.026259 2011;134(pt 6):1765–1776. doi: 10.1093/brain/awr100
17. Chatterjee A, Somayaji NR, Kabakis IM. Artificial intelligence detection 33. Stier N, Vincent N, Liebeskind D, Scalzo F. Deep learning of tissue fate features
of cerebrovascular large vessel occlusion - nine month, 650 patient in acute ischemic stroke. Proceedings (IEEE Int Conf Bioinformatics Biomed).
evaluation of the diagnostic accuracy and performance of the viz.Ai 2015;2015:1316–1321. doi: 10.1109/BIBM.2015.7359869
lvo algorithm. Proc AHA Stroke. 2019;50:AWMP16-AWMP16. 34. Nielsen A, Hansen MB, Tietze A, Mouridsen K. Prediction of tissue
18. Straka M, Albers GW, Bammer R. Real-time diffusion-perfusion out-come and assessment of treatment effect in acute ischemic stroke
mismatch analysis in acute stroke. J Magn Reson Imaging. using deep learning. Stroke. 2018;49:1394–1401. doi: 10.1161/
2010;32:1024–1037. doi: 10.1002/jmri.22338 STROKEAHA.117.019740
19. Albers GW, Thijs VN, Wechsler L, Kemp S, Schlaug G, Skalabrin E, 35. Yu Y, Xie Y, Thamm T, Gong E, Quyang J, Huang C, Christensen S, Marks
Bammer R, Kakuda W, Lansberg MG, Shuaib A, et al; DEFUSE MP, Lansberg MG, Albers GW, et al. Use of deep learning to predict final
Investigators. Mag-netic resonance imaging profiles predict clinical ischemic stroke lesions from initial magnetic resonance imaging. JAMA
response to early reper-fusion: the diffusion and perfusion imaging Net-work Open. 2020. doi: 10.1001/jamanetworkopen.2020.0772.
Downloaded from http://ahajournals.org by on November 16, 2020
evaluation for understanding stroke evolution (DEFUSE) study. Ann 36. Pérez Malla CU, Valdés Hernández MDC, Rachmadi MF, Komura T.
Neurol. 2006;60:508–517. doi: 10.1002/ana.20976 Evalu-ation of enhanced learning techniques for segmenting
20. Olivot JM, Mlynash M, Thijs VN, Kemp S, Lansberg MG, Wechsler L, ischaemic stroke lesions in brain magnetic resonance perfusion
Bammer R, Marks MP, Albers GW. Optimal Tmax threshold for images using a convo-lutional neural network scheme. Front
predict-ing penumbral tissue in acute stroke. Stroke. 2009;40:469– Neuroinform. 2019;13:33. doi: 10.3389/fninf.2019.00033
475. doi: 10.1161/STROKEAHA.108.526954 37. Yu Y, Guo D, Lou M, Liebeskind D, Scalzo F. Prediction of hemorrhagic
21. Huang X, Kalladka D, Cheripelli BK, Moreton FC, Muir KW. The trans-formation severity in acute stroke from source perfusion MRI. IEEE
impact of CT perfusion threshold on predicted viable and nonviable Trans Biomed Eng. 2018;65:2058–2065. doi: 10.1109/TBME.2017.2783241
tissue vol-umes in acute ischemic stroke. J Neuroimaging. 38. Yu Y, Guo D, Lou M, Liebeskind D, Scalzo F. The prediction of the hemor-rhagic
2017;27:602–606. doi: 10.1111/jon.12442 transformation locations after reperfusion therapy in acute stroke patients: a
22. Austein F, Riedel C, Kerby T, Meyne J, Binder A, Lindner T, Huhndorf perfusion study using deep learning. Neurology. 2018;90:P3.212.
M, Wodarg F, Jansen O. Comparison of perfusion CT software to 39. Bentley P, Ganesalingam J, Carlton Jones AL, Mahady K, Epton S,
predict the final infarct volume after thrombectomy. Stroke. Rinne P, Sharma P, Halse O, Mehta A, Rueckert D. Prediction of
2016;47:2311–2317. doi: 10.1161/STROKEAHA.116.013147 stroke thrombolysis outcome using CT brain machine learning.
23. Chen L, Bentley P, Rueckert D. Fully automatic acute ischemic lesion Neuroimage Clin. 2014;4:635– 640. doi: 10.1016/j.nicl.2014.02.003
seg-mentation in DWI using convolutional neural networks. 40. van Os HJA, Ramos LA, Hilbert A, van Leeuwen M, van Walderveen MAA,
Neuroimage Clin. 2017;15:633–643. doi: 10.1016/j.nicl.2017.06.016 Kruyt ND, Dippel DWJ, Steyerberg EW, van der Schaaf IC, Lingsma HF, et
24. Ronneberger O., Fischer P., Brox T. (2015) U-Net: Convolutional Networks for al. Predicting outcome of endovascular treatment for acute ischemic stroke:
Biomedical Image Segmentation. In: Navab N., Hornegger J., Wells potential value of machine learning algorithms. Front Neurol. 2018;9:784.
W., Frangi A. (eds) Medical Image Computing and Computer-Assisted 41. Xie Y, Jiang B, Gong E, Li Y, Zhu G, Michel P, Wintermark M, Zaharchuk
Intervention - MICCAI 2015. MICCAI 2015. Lecture Notes in G. Use of gradient boosting machine learning to predict patient outcome in
Computer Science, vol 9351. Springer, Cham. acute ischemic stroke on the basis of imaging, demographic, and clinical
25. Song T. 3D multi-scale U-net with atrous convolution for ischemic stroke information. Am J Roentgenol. 2019;212:44–51.
lesion segmentation. Proc MICCAI ISLES 2018 Challenge. 2018. 42. Bacchi S, Zerner T, Oakden-Rayner L, Kleinig T, Patel S, Jannes J.
26. Winzeck S, Hakim A, McKinley R, Pinto JAADSR, Alves V, Silva C, Deep Learning in the prediction of ischaemic stroke thrombolysis func-
Pisov M, Krivov E, Belyaev M, Monteiro M, et al. ISLES 2016 and tional outcomes: a pilot study. Acad Radiol. 2020;27:e19–e23. doi:
2017-Benchmarking ischemic stroke lesion outcome prediction based 10.1016/j.acra.2019.03.015