Abstract
Background and purpose:
Computed Tomography Perfusion imaging (CTP) allows estimation of tissue status in patients with acute ischemic stroke. We aimed to improve prediction of the final infarct and individual infarct growth rates using a deep learning approach.
Methods:
We trained a deep neural network to predict the final infarct volume in acute stroke patients presenting with large vessel occlusions based on the native CTP images, time to reperfusion and reperfusion status in a derivation cohort (MR CLEAN trial). The model was internally validated in a five-fold cross-validation and externally in an independent dataset (CRISP study). We calculated the mean absolute difference (MAD) between the predictions of the deep learning model and the final infarct volume versus the MAD between CTP processing by RAPID software and the final infarct volume. Next, we determined infarct growth rates for every patient.
Results:
We included 127 patients from the MR CLEAN (derivation) and 101 patients of the CRISP study (validation). The deep learning model improved final infarct volume prediction compared to the RAPID software in both the derivation, MAD 34.5 vs 52.4ml, and validation cohort, 41.2 vs 52.4 ml, (p < 0.01). We obtained individual infarct growth rates enabling the estimation of final infarct volume based on time and grade of reperfusion.
Conclusions:
We validated a deep learning-based method which improved final infarct volume estimations compared to classic CTP processing. In addition, the deep learning model predicted individual infarct growth rates which could enable the introduction of tissue clocks during the management of acute stroke.
Graphical Abstract
Introduction
Acute ischemic stroke treatment aims at salvaging the hypoperfused tissue by recanalizing the occluded artery by intravenous thrombolysis or endovascular treatment (EVT). For years, expected time of symptom onset to treatment determined eligibility for these therapeutic regimens 1–3. Recent trials have provided evidence for efficacy of EVT after imaging-based selection in patients presenting outside of the conventional time windows of stroke onset 4,5.
Perfusion imaging, most commonly CT perfusion (CTP), can aid in identifying the potentially salvageable tissue (i.e. penumbra) calculated by subtracting the irreversibly damaged ischemic core from the hypoperfused lesion volume 6–8. The currently clinically available software programs calculate different perfusion parameters by using a deconvolution technique 9,10. In the clinical trials DEFUSE 3 (Endovascular Therapy Following Imaging Evaluation for Ischemic Stroke), SWIFT-PRIME (Stent-Retriever Thrombectomy after intravenous t-PA vs. t-PA alone in stroke), DAWN (Thrombectomy 6 to 24 hours after Stroke with a Mismatch between Deficit and Infarct) and EXTEND (Thrombolysis Guided by Perfusion Imaging up to 9 Hours after Onset of Stroke), a relative cerebral blood flow (rCBF) less than 30% determined the ischemic core volume and a delay to the maximum of the residue function (Tmax) exceeding 6s identified the critically hypoperfused tissue 4,5,11,12. Patients were selected for inclusion in these trials based on a predefined mismatch pattern of core and perfusion lesion estimates 13,14. However, the calculated maps depend on the software program used and can still require human input to distinguish artifacts from real perfusion deficits 10. In addition, this deconvolution technique is inherently sensitive to noise and therefore improving the current perfusion algorithms or even use deconvolution-free summary parameters is the aim of current research 15–17. Machine learning is an alternative to estimate tissue fate in ischemic stroke patients 18–20. The current classification of tissue into penumbra and core relate to the moment of scanning and does not project the rate of infarct evolution which can be useful in transfer settings. The rate of infarct growth varies greatly between patients and is among other things dependent on the collateral circulation 21,22. The hypoperfusion intensity ratio (HIR), which ranges between 0–1, obtained at baseline with perfusion imaging correlated with actual infarct growth rates and a HIR-value < 0.5 suggested infarct growth rates below 5ml/h 23. A better insight in the individual growth prediction could influence logistics regarding both transfer of patients as well as time metrics in comprehensive stroke centers.
We previously developed a deconvolution free algorithm based on deep learning to predict tissue status directly from CTP source images 24. The aim of this study was to compare final infarct volume estimation accuracy between a deep neural network and a classic deconvolution/thresholding analysis. Additionally, we aimed to provide individual lesion growth rates in patients with a large vessel occlusion and correlate these estimations to the HIR.
Methods
The data that support the findings of this study are available from the corresponding author upon reasonable request.
The local institutional review boards from all participating institutions approved the original studies.
Deep learning model
We trained a deep neural network to predict final infarct size in acute ischemic stroke patients with large vessel occlusions in the anterior circulation 2. We refer to the original paper for the full technical details of the deep learning model (DL model) 24. To summarize, the network was trained based on the CTP source images from patients included in ‘Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands’ (MRCLEAN; the derivation cohort) in a voxelwise classification approach. Four clinical parameters were added to the DL model: time of symptom onset to imaging, time to recanalization, modified Thrombolysis In Cerebral Infarction (mTICI) scores at the end of the procedure and the persistence of the arterial occlusion on CT angiography (CTA) at 24h. The output of the DL model provides the probabilities of infarction per voxel. The predicted final infarct volumes are calculated by taking the sum of all probabilities and multiply this result with the volume per voxel (Figure 1). Upon request we added the MI-CLAIM checklist to the supplementary data as a quality check.25
Figure 1.
Example of the probabilistic prediction made by the deep learning model using the native perfusion images and clinical parameters. The red voxels represent the probabilistic prediction of the final infarct by the deep learning model and the blue line depicts the actual final infarct.
Validation of the final infarct volume prediction
We compared the performance of the DL model in predicting the final infarct volume in two cohorts of patients to a clinically validated CT perfusion software package (i.e. RAPID) which uses a deconvolution/thresholding approach 9. The DL model was internally validated in a five-fold cross-validation in the derivation cohort (MRCLEAN) 2. The cohort was randomly split in 5 subcohorts and the model was trained five times on the subjects of four subcohorts and evaluated based on the predictions for the subjects in the other subcohort 24. As a result, we obtained predictions for each subject by a model that was not trained on that subject. Here, the DL model was validated externally in an independent dataset from the ‘Computed tomographic perfusion to predict response to recanalization in ischemic stroke’ (CRISP) study 26.
We used the manually delineated lesions (from the original study) at follow up CT scan on day 5 (or when not available at 24h) in the MR CLEAN study to define the gold standard for the final infarct volumes. For the CRISP study these final volumes had been determined by the core lab based on the Fluid Attenuated Inversion Recovery (FLAIR) images of day 5 27,28. The same slice coverage as for the initial CTP scan was considered for the calculation of the final infarct volume.
The DL model can predict the final infarct volume for any hypothetical level of achieved reperfusion. Inherently the deconvolution/thresholding approach can predict the final infarct only in the absence of reperfusion (baseline hypoperfused lesion volumes serves as a proxy for the final infarct volume) or if complete reperfusion is achieved (where the baseline infarct core will equal the final infarct volume). Poor and intermediate reperfusion status will result in an infarct volume which cannot be estimated by the deconvolution/thresholding approach. To address this issue, we excluded patients with intermediate reperfusion when comparing baseline predictions vs actual final infarct volumes for both models. Successful reperfusion was defined as a mTICI scale 2b or 3 in patients who underwent EVT; poor (mTICI 1) or intermediate reperfusion (mTICI 2a); and no reperfusion in patients with a final mTICI of 0 and in addition patients without EVT in whom CTA at 24h showed a persistent large vessel occlusion. In patients with successful reperfusion the core volume at baseline, defined as the area with a reduction of the relative cerebral blood flow (rCBF) below 30%, served as the prediction of the final infarct volume. In patients without reperfusion the hypoperfused lesion volume, delineated by Tmax maps with a delay of more than 6s, was regarded as the predicted final infarct volume 5,7,29. These thresholds have been used in many clinical trials for patient selection for reperfusion therapy and are implemented by RAPID software as the default output maps 5,11. We visually inspected all the perfusion maps and manually corrected the arterial input function (AIF) when automatic selection failed.
Evaluating the ischemic core with CTP remains challenging and although the threshold of a rCBF<30% is clinically relevant and validated, the mean absolute volumetric differences between the CTP vs acute diffusion weighted imaging (DWI) core volume is the smallest when using a threshold of 38% 7. Therefore, in an exploratory analysis, we calculated final infarct volumes in patients with reperfusion based on the rCBF threshold of 38%.
Growth prediction and association with collateral grades in CRISP
We computed infarct volumes as predicted by the deep learning model at different time points after imaging and with various grades of reperfusion in the validation sample of the CRISP study. Based on these data, we calculated the mean lesion growth rates per 60 minutes for patients in a time window between imaging and reperfusion of 30–180 minutes with simulated mTICI scores. We only show results for clinically relevant mTICI scores of 2a, 2b or 3.
We obtained the hypoperfusion intensity ratio (HIR) as a marker of collateral status, which is defined as the proportion of Tmax > 6s volume with Tmax > 10s 30. HIR correlates well with the quality of collateral circulation determined by angiography 30. We investigated the association between the HIR and predicted lesion growth in patients with a target mismatch profile similar as has been reported for actual growth in patients during transfer to a comprehensive stroke center 23. This target mismatch profile was defined according to DEFUSE 3 criteria as a core volume below 70 ml, an absolute difference in Tmax > 6s minus core volume of at least 15 ml and a ratio of the Tmax > 6s lesion over the core volume of 1.8 or higher 5.
Statistical analysis
Characteristics of the two study populations were compared with a Chi-Square test for categorical variables and with the Mann-Whitney U test for continuous variables.
We calculated mean absolute volumetric differences (MAD) between the predicted and the gold standard final infarct volumes. These volume errors were compared between the deep learning model vs RAPID software by the Wilcoxon signed-rank test.
Receiver operating characteristic (ROC) curves are depicted for HIR values to discriminate between slow and fast predicted rates of lesion growth with optimal sensitivity and specificity. Following the definitions of a previous publication, the cut off for slow vs fast infarct growth was set to 5ml/h 23. The optimal HIR threshold to separate slow and fast lesion growth was calculated by the Youden index. In addition, we compared the difference in median lesion growth between patients with good and poor collateral circulation with the Mann-Whitney U test.
R software was used for statistical analysis 31. P-values below 0.05 were considered significant.
Results
Demographics
We included 127 of the 500 patients randomized in the MR CLEAN trial. Reasons for exclusion were: absence of baseline CTP (231), insufficient quality of baseline CTP (42), incomplete information on reperfusion status (59) and lack of follow up non-contrast CT (NCCT) (41). For the external validation we could analyze 104 of the 190 patients enrolled in CRISP. Most patients (84) were excluded since follow up FLAIR at day 5 was unavailable, and in 2 patients the CTP data was of insufficient quality (Online figure I). We manually selected the AIF in 4 patients. Clinical - and imaging characteristics of the included patients are listed in Table 1. Patients from the CRISP dataset were older, had smaller baseline core volumes, presented later after symptom onset for CTP imaging, had shorter imaging to reperfusion time windows and higher reperfusion rates. CRISP was a cohort study of patients undergoing EVT and the MR CLEAN trial was a randomized controlled trial in which half of the patients did not receive EVT. This at least partially explains the lower rates of successful reperfusion in patients included from the MR CLEAN trial.
Table 1.
Demographic data
| MRCLEAN (n= 127) | CRISP (n= 104) | p-value | |
|---|---|---|---|
| Age | 63 (52–73) | 68 (57–76) | 0.01 |
| Gender (f) | 63 (50%) | 52 (50%) | 1 |
| NIHSS D0 | 17 (14–21) | 18 (12–21) | 0.36 |
| Time of symptom onset to CTP (min) | 171 (110–233) | 273 (154–406) | < 0.01 |
| Time of CTP to recanalization (min) | 163 (136–217) | 142 (106–174) | < 0.01 |
| Core volume RAPID (ml) | 14.1 (0–39.9) | 7.6 (0–20.3) | 0.01 |
| mTICI score 2b – 3 * | 48 (38%) | 87 (84%) | < 0.01 |
| Final infarct volume (ml) | 51 (21.3–112.5) | 52.0 (25.7–109.0) | 0.80 |
Data are median (IQR) or n (%).
NIHSS = National Institute of Health Stroke Scale; CTP = Computed Tomography Perfusion; f = Female; mTICI = modified Thrombolysis In Cerebral Infarction;
For MR CLEAN in 48 of 75 patients (64%) randomized to the interventional arm the mTICI score was 2b-3.
Validation of final infarct volume prediction
For the internal validation of the MR CLEAN study we could include 48 patients (36%) with successful and 60 patients (44%) without reperfusion to assess the baseline prediction of the final infarct size in comparison to RAPID software. When comparing the final infarct estimation in these 108 patients based on the RAPID analysis of the baseline CTP vs the actual infarct volume the MAD was 52.4 ml (SD 49.8). The DL model provided a better estimate of the final infarct with a MAD of 34.5 ml (SD 29.4) (p < 0.01) (Figure 2A, Online Table I).
Figure 2.
A. Internal validation: Scatterplots for predicted volumes (ml) for the MRCLEAN population (n = 108).
B. External validation: Scatterplots for predicted volumes (ml) for the CRISP population (n = 92). Black line depicts the identity line.
In 19 patients with intermediate reperfusion from the MR CLEAN study the MAD for the DL model vs the actual final infarct volume was 36.7 ml (SD 38.3). This result didn’t differ from the group with complete or no reperfusion (p=0.64).
In the external validation of the CRISP study we documented successful reperfusion in 87 patients (84%) and no reperfusion in 5 (5%). In this independent cohort of 92 patients the MAD between baseline and actual final infarct volume was also smaller (41.2 ml (SD 55.4)) for the DL model compared to the prediction by the RAPID software (52.4 ml (SD 58.3); p < 0.01) (Figure 2B, Online Table I). Bland-Altman plots for the internal and external validation sets are depicted in Online figure II.
In an exploratory analysis of the CRISP dataset we applied a different rCBF threshold of 38% to calculate the core volume (i.e. the predicted final infarct volume in reperfused patients) with the RAPID software. This resulted in a lower MAD (42.2 ml (SD 54.5)) for the calculations by the automated RAPID software. At this threshold both the DL model and RAPID software performed similarly in predicting the final infarct volume (p-value for difference in MAD = 0.3).
In the CRISP dataset, 12 patients presented with intermediate reperfusion and the MAD for the DL model vs the final infarct volume was 58.5 ml (SD 63.5). This difference was comparable to the group with complete or no reperfusion (p=0.46).
Individual growth modelling
Based on the information of the baseline CTP, the deep learning model can provide infarct growth rates in individual patients. With the input of a particular reperfusion status (mTICI 2a, 2b or 3) at a chosen time point after imaging, the model is able to estimate the corresponding final infarct volume. The evolution of infarct growth can be presented in a graph and transformed in a movie to provide clinicians with an interactive visualization. In Figure 3 and 4 we show examples of such growth predictions for different patients (corresponding movies in Online video I–II).
Figure 3.
A. Patient with a mean infarct growth of 18.3 ml/h. The final infarct volume was 104 ml. Recanalization was performed 131 min after CT perfusion with a mTICI = 2b.
B. Patient with a mean infarct growth of 2.3 ml/h. The final infarct volume was 10.8 ml. Recanalization was performed 101 min after CT perfusion with a mTICI = 3.
Figure 4.
Final infarct volumes for complete reperfusion as predicted by the deep learning model are illustrated in red at different time points. Patient A had a mean infarct growth of 9.6 ml/h and patient B of 12.8 ml/h. Corresponding movies can be found in Supplementary Movies.
We calculated the HIR and projected mean lesion growth rates per hour for patients who presented with a target mismatch profile (n=98, 94%). Figure 5A depicts the ROC curve of the HIR as a predictor of patients with fast (> 5ml/h) vs slow infarct growth rates (≤ 5ml/h). The Youden index revealed an optimal threshold of 0.36 to discriminate between these two growth patterns. In patients with a HIR < 0.36 the median lesion growth rate was lower (2.7 ml/h) compared to patients with a HIR ≥ 0.36, in whom the median growth rate was 8.5 ml/h (p < 0.01) (Figure 5B). In Online Figure III the scatterplot between HIR and predicted infarct growth rate is depicted.
Figure 5.
A. Receiver operating characteristics curve for hypoperfusion intensity ratio (HIR) to predict lesion growth of 5ml/h in patients with the target mismatch profile in CRISP (n = 98). Youden index = 0.36. AUC = Area under the curve.
B. Boxplots for the predicted infarct growth (ml/h) depending on a hypoperfusion intensity ratio (HIR) < 0.36 (n = 36) vs ≥ 0.36 (n= 62) (p <0.01).
Discussion
We present a deep learning model to predict final infarct volumes at baseline based on CTP images and reperfusion status in patients with an acute ischemic stroke caused by a large vessel occlusion. The model improved estimations of final infarct volumes in two independent cohorts compared to a clinically validated automated software. This deep learning model provided infarct growth rates for individual patients and we replicated a correlation between the HIR and these growth rates.
The randomized clinical trials of EVT in the extended time window included patients in whom tissue status was known based on CTP or MR imaging, underscoring the importance of more advance neuro-imaging compared to only NCCT and CTA 4,5. Although software packages to identify core and penumbra are clinically validated, limitations of the classic deconvolution technique are well described in the literature and current research focuses on more advanced learning approaches. Previously published deep learning models made use of the different perfusion parameters obtained by deconvolution as input parameters or the output derived from deconvolution as ground truth 18,32–35. Our model is completely free of deconvolution by using the source perfusion images as input and the final infarct as ground truth, which is a great advantage. In previous research we showed that source perfusion scans as input parameters outperformed perfusion parameters derived from deconvolution in predicting final infarct volume 24.
CTP imaging reveals infarcted and hypoperfused brain tissue based on hemodynamic changes resulting from the occlusion of an artery. A decrease in the CBF has been validated as the most accurate prediction of the ischemic core with a threshold of a rCBF< 38%. Calculations of core volumes based on this threshold will occasionally overestimate the infarcted tissue, potentially resulting in the exclusion of candidates for EVT. In clinical practice preventing overcalling of the infarcted tissue is favored over accuracy and therefore the more strict threshold of a rCBF<30% was introduced in the RAPID software to determine core 7. The perfusion lesion volume (i.e. the area in the brain which is critically hypoperfused) is derived from the Tmax parameter revealing delays above 6s 36. In patients with successful reperfusion the ischemic core obtained at baseline frequently equals the final infarct volume and in patients without reperfusion the estimated perfusion lesion volume acts as a prediction for the final infarct size 12. Any intermediate reperfusion status will result in stroke sizes between the baseline core and perfusion lesion volumes, but accurate prediction is unobtainable from conventional CTP analyses. Therefore, the comparison of the MAD between volumes acquired by conventional CTP-based automated software vs the actual final infarct could only be determined in the group of patients with and without reperfusion (i.e. not in patients with intermediate reperfusion).
The deep learning model produced lower MADs in both the MR CLEAN and CRISP study compared to the RAPID software, one of the most commonly used perfusion software programs that is based on a classic deconvolution/thresholding technique. One of the advantages of the deep learning model might be the integration of the timing of the imaging related to the onset of symptoms and to the timing of recanalization. However, we do admit that the mean absolute differences remain quite large which is a known challenge in estimating infarct volumes at baseline, likely due to the inherent limitation of the current perfusion imaging.17
Improving imaging acquisition techniques and developing new sequences could be the focus of future research goals.
We performed an exploratory analysis comparing the DL to the RAPID analysis with the rCBF<38% threshold. This analysis underscored the previous findings of a more accurate prediction with this threshold since the MAD was no longer different between the two methods.
In the group with intermediate reperfusion we could only investigate the performance of the DL model which revealed similar volumetrical differences to the group with complete or no reperfusion.
An important advantage of the deep learning approach is the possibility to predict the final infarct volume in different scenarios for both time to recanalization and mTICI scores. In addition, we created individual growth rates which can be visualized in graphs and movies displaying the evolution of the infarct. This tissue clock can guide physicians in several decisions like the indication to repeat neuro-imaging in transfer patients and may encourage teams in comprehensive stroke centers during EVT procedures to improve time metrics. Models predicting infarct growth rates can also be of importance in future trials with neuroprotective agents. The most optimal patients will be those with substantial infarct growth in the therapeutic time window between imaging and reperfusion. Patients with slow infarct growth rates are not likely good candidates for these neuroprotective treatments since considerable infarct growth between imaging and reperfusion is not expected 37,38. In a previous study, collateral status, as assessed by the HIR at baseline, correlated with actual infarct growth rates calculated based on two neuro-imaging studies before and after transfer of patients to a comprehensive stroke center 23. With our model we identified a similar association between the HIR and the predicted lesion growth rate, which may represent an indirect confirmation of the growth prediction model. The suggested threshold of 0.36 closely approximates the threshold of 0.4 as predictor of collateral status defined on digital subtraction angiography and infarct growth in the original paper that introduced the HIR 30. This analysis was limited to patients with a target mismatch profile, which might have led to an exclusion of the fast growers. Therefore, validation in a cohort of consecutive stroke patients with large vessel occlusions would be of interest to increase the heterogeneity in perfusion profiles.
A general limitation of a deep learning model is the difficult interpretation of the origin of the individual results. One can only interpret the probabilities of the final infarct prediction, but not how they are exactly calculated by the DL model. Furthermore, our results are only applicable to patients with similar characteristics as in the described study population. Further research is warranted in other cohorts as for instance in patients with posterior stroke.
Another limitation of our study is the lack of a uniform method to quantify the final infarct volume. In the derivation cohort we determined the final infarct on NCCT imaging 5 days after stroke onset and if unavailable we used the 24h NCCT. The latter may be suboptimal since in a substudy of MR CLEAN some infarct growth occurred after 24h 27. Therefore previously we evaluated the two methods as measurement of the final infarct volumes and found no difference in performance of the DL model 24. In the validation cohort we delineated final infarct volumes on the day 5 FLAIR to enable a more accurate volumetric approach, since early MRI can result in underestimation especially in patients who do not reperfuse 28,39. Another limitation is the lack of a gold standard to validate the calculated lesion growth rates. In MR CLEAN and CRISP, patients did not undergo two neuro-imaging studies before reperfusion therapy hampering measurements of actual growth rates. Furthermore, hemorrhagic transformation may also contribute to infarct growth. The similar association with the HIR and our predicted infarct growth rates compared to the actual infarct growth found in a smaller study could serve as an indirect validation of our model 23.
Conclusion
We validated an innovative deep learning approach to predict final infarct volumes directly from the CT source perfusion images in a large independent dataset of patients with acute large vessel ischemic strokes. The model can derive individual infarct growth rates which could become part of clinical routine in both primary and comprehensive stroke centers.
Supplementary Material
Acknowledgments
Funding Sources
The MR CLEAN trial was partly funded by the Dutch Heart Foundation and by unrestricted grants from AngioCare BV, Medtronic/Covidien/EV3, Medac Gmbh/Lamepro, Penumbra Inc., Stryker and Top Medical/Concentric.
Disclosures
Dr. Marquering reports co-founder and shareholder of Nico-lab outside the submitted work.
Dr. Roos reports stock ownership of Nico-Lab outside the submitted work. Dr. Majoie reports grants from CVON/Dutch Heart Foundation, grants from European Commission, grants from Dutch Health Evaluation Program, grants from Stryker, and grants from TWIN Foundation (Toegepast Wetenschappelijk Instituut voor Neuromodulatie) outside the submitted work; and is shareholder of Nico-lab. Dr. van der Lugt reports grants from Angiocare BV, grants from Covidien / EV3, grants from Medac GmbH / Lamepro, grants from Stryker, grants from Penumbra, and grants from Dutch Heart Foundation during the conduct of the study; grants from stryker, grants from Medtronic, grants from penumbra, grants from cerenovus, and grants from thrombolytic science Inc outside the submitted work. Dr. Lansberg reports grants from NIH during the conduct of the study. Dr. Albers reports equity interest in iSchemaView, owner of RAPID software which is used in this study.
Dr. Christensen reports equity interest in iSchemaView, owner of RAPID software which is used in this study.
Dr. Dippel reports grants from Dutch Heart Foundation, grants from AngioCare BV, grants from Covidien/EV3®, grants from MEDAC Gmbh/LAMEPRO, grants from Penumbra Inc., grants from Top Medical/Concentric, grants from Stryker, during the conduct of the study.
Dr van Zwam reports personal fees from Cerenovus and personal fees from Stryker during the conduct of the study.
Dr. Robben reports grants from Flanders Innovation and Entrepreneurship (VLAIO) during the conduct of the study; in addition, Dr Robben has a patent to Patent for automatic AIF selection pending.
Dr. Lemmens is a senior clinical investigator for Fund for Scientific Research-Flanders (FWO: 1841918N), reports a grant from FWO Flanders (G049620N) and institutional support for consultancy by iSchemaView.
Non-standard Abbreviations and Acronyms
- AIF
Arterial input function
- CTA
Computed Tomography Angiography imaging
- CTP
Computed Tomography Perfusion imaging
- DL
Deep learning
- DWI
Diffusion weighted imaging
- EVT
Endovascular treatment
- FLAIR
Fluid-attenuated inversion recovery
- HIR
Hypoperfusion intensity ratio
- MAD
Mean absolute volumetric differences
- NCCT
Non-contrast computed tomography
- mTICI
Modified Thrombolysis In Cerebral Infarction
- rCBF
Relative cerebral blood flow
- ROC
Receiver operating characteristic
References
- 1.Hacke W, Kaste M, Bluhmki E, Brozman M, Dávalos A, Guidetti D, Larrue V, Lees KR, Medeghri Z, Machnig T, et al. Thrombolysis with alteplase 3 to 4.5 hours after acute ischemic stroke. N. Engl. J. Med 2008;359:1317–1329. [DOI] [PubMed] [Google Scholar]
- 2.Berkhemer O, Fransen P, Beumer D, van der Berg L, Lingsma H, Yoo A, Schonewille W, Vos J, Nederkoorn P, Wermer M, et al. A Randomized Trial of Intraarterial Treatment for Acute Ischemic Stroke. New Engl. J. Med 2014;372:11–20. [DOI] [PubMed] [Google Scholar]
- 3.Saver JL, Goyal M, van der Lugt A, Menon BK, Majoie CBLM, Dippel DW, Campbell BC, Nogueira RG, Demchuk AM, Tomasello A, et al. Time to Treatment With Endovascular Thrombectomy and Outcomes From Ischemic Stroke: A Meta-analysis. Jama. 2016;316:1279–1288. [DOI] [PubMed] [Google Scholar]
- 4.Nogueira RG, Jadhav AP, Haussen DC, Bonafe A, Budzik RF, Bhuva P, Yavagal DR, Ribo M, Cognard C, Hanel RA, et al. Thrombectomy 6 to 24 hours after stroke with a mismatch between deficit and infarct. N. Engl. J. Med 2018;378:11–21. [DOI] [PubMed] [Google Scholar]
- 5.Albers GW, Marks MP, Kemp S, Christensen S, Tsai JP, Ortega-Gutierrez S, McTaggart RA, Torbey MT, Kim-Tenser M, Leslie-Mazwi T, et al. Thrombectomy for stroke at 6 to 16 hours with selection by perfusion imaging. N. Engl. J. Med 2018;378:708–718. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Purushotham A, Campbell BC V, Straka M, Mlynash M, Olivot J-M, Bammer R, Kemp SM, Albers GW, Lansberg MG. Apparent diffusion coefficient threshold for delineation of ischemic core. Int. J. Stroke 2013;1–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Cereda CW, Christensen S, Campbell BC, Mishra NK, Mlynash M, Levi C, Straka M, Wintermark M, Bammer R, Albers GW, et al. A benchmarking tool to evaluate computer tomography perfusion infarct core predictions against a DWI standard. J. Cereb. Blood Flow Metab 2016;36:1780–1789. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Campbell BCV, Christensen S, Levi CR, Desmond PM, Donnan G a, Davis SM, Parsons MW. Cerebral blood flow is the optimal CT perfusion parameter for assessing infarct core. Stroke. 2011;42:3435–40. [DOI] [PubMed] [Google Scholar]
- 9.Straka M, Albers GW, Bammer R. Real-time diffusion-perfusion mismatch analysis in acute stroke. J. Magn. Reson. Imaging 2010;32:1024–1037. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Austein F, Riedel C, Kerby T, Meyne J, Binder A, Lindner T, Huhndorf M, Wodarg F, Jansen O. Comparison of Perfusion CT Software to Predict the Final Infarct Volume after Thrombectomy. Stroke. 2016;47:2311–2317. [DOI] [PubMed] [Google Scholar]
- 11.Ma H, Campbell BCV, Parsons MW, Churilov L, Levi CR, Hsu C, Kleinig TJ, Wijeratne T, Curtze S, Dewey HM, et al. Thrombolysis Guided by Perfusion Imaging up to 9 Hours after Onset of Stroke. N. Engl. J. Med 2019;380:1795–1803. [DOI] [PubMed] [Google Scholar]
- 12.Saver J, Goyal M, Bonafe A, Diener H, Levy E, Pereira V. Stent-Retriever Thrombectomy after Intravenous t-PA vs. t-PA Alone in Stroke. N. Engl. J. Med 2015;372:2285–2295. [DOI] [PubMed] [Google Scholar]
- 13.Albers GW, Thijs VN, Wechsler L, Kemp S, Schlaug G, Skalabrin E, Bammer R, Kakuda W, Lansberg MG, Shuaib A, et al. Magnetic Resonance Imaging Profiles Predict Clinical Response to Early Reperfusion : The Diffusion and Perfusion Imaging Evaluation for Understanding Stroke Evolution ( DEFUSE ) Study. Ann. Neurol 2006;60:508–517. [DOI] [PubMed] [Google Scholar]
- 14.Ma H, Wright P, Allport L, Phan TG, Churilov L, Ly J, Zavala J a, Arakawa S, Campbell B, Davis SM, et al. Salvage of the PWI/DWI mismatch up to 48 h from stroke onset leads to favorable clinical outcome. Int. J. Stroke 2014;1–6. [DOI] [PubMed] [Google Scholar]
- 15.Meijs M, Christensen S, Lansberg MG, Albers GW, Calamante F. Analysis of perfusion MRI in stroke: To deconvolve, or not to deconvolve. Magn. Reson. Med 2016;76:1282–1290. [DOI] [PubMed] [Google Scholar]
- 16.Wouters A, Christensen S, Straka M, Mlynash M, Liggins J, Bammer R, Thijs V, Lemmens R, Albers GW, Lansberg MG. A comparison of relative time to peak and tmax for mismatch-based patient selection. Front. Neurol 2017;8:1–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Bouslama M, Ravindran K, Harston G, Rodrigues GM, Pisani L, Haussen DC, Frankel MR, Nogueira RG. Noncontrast Computed Tomography e-Stroke Infarct Volume Is Similar to RAPID Computed Tomography Perfusion in Estimating Postreperfusion Infarct Volumes. Stroke. 2021;52:634–641. [DOI] [PubMed] [Google Scholar]
- 18.Yu Y, Xie Y, Thamm T, Gong E, Ouyang J, Huang C, Christensen S, Marks MP, Lansberg MG, Albers GW, et al. Use of Deep Learning to Predict Final Ischemic Stroke Lesions From Initial Magnetic Resonance Imaging. JAMA Netw. Open 2020;3:e200772. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Nielsen A, Hansen MB, Tietze A, Mouridsen K. Prediction of tissue outcome and assessment of treatment effect in acute ischemic stroke using deep learning. Stroke. 2018;49:1394–1401. [DOI] [PubMed] [Google Scholar]
- 20.Kemmling A, Flottmann F, Forkert ND, Minnerup J, Heindel W, Thomalla G, Eckert B, Knauth M, Psychogios M, Langner S, et al. Multivariate dynamic prediction of ischemic infarction and tissue salvage as a function of time and degree of recanalization. J. Cereb. Blood Flow Metab 2015;35:1397–1405. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Wheeler HM, Mlynash M, Inoue M, Tipirnini A, Liggins J, Bammer R, Lansberg MG, Kemp S, Zaharchuk G, Straka M, et al. The growth rate of early DWI lesions is highly variable and associated with penumbral salvage and clinical outcomes following endovascular reperfusion. Int. J. Stroke 2015;10:723–729. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.De Havenon A, Mlynash M, Kim-Tenser MA, Lansberg MG, Leslie-Mazwi T, Christensen S, McTaggart RA, Alexander M, Albers G, Broderick J, et al. Results from DEFUSE 3: Good collaterals are associated with reduced ischemic core growth but not neurologic outcome. Stroke. 2019;50:632–638. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Guenego A, Mlynash M, Christensen S, Kemp S, Heit JJ, Lansberg MG, Albers GW. Hypoperfusion ratio predicts infarct growth during transfer for thrombectomy. Ann. Neurol 2018;84:616–620. [DOI] [PubMed] [Google Scholar]
- 24.Robben D, Boers AMM, Marquering HA, Langezaal LLCM, Roos YBWEM, van Oostenbrugge RJ, van Zwam WH, Dippel DWJ, Majoie CBLM, van der Lugt A, et al. Prediction of final infarct volume from native CT perfusion and treatment parameters using deep learning. Med. Image Anal 2020;59. [DOI] [PubMed] [Google Scholar]
- 25.Norgeot B, Quer G, Beaulieu-Jones BK, Torkamani A, Dias R, Gianfrancesco M, Arnaout R, Kohane IS, Saria S, Topol E, et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat. Med 2020;26:1320–1324. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Lansberg MG, Christensen S, Kemp S, Mlynash M, Mishra N, Federau C, Tsai JP, Kim S, Nogueria RG, Jovin T, et al. Computed tomographic perfusion to Predict Response to Recanalization in ischemic stroke. Ann. Neurol 2017;81:849–856. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Bucker A, Boers AM, Bot JCJ, Berkhemer OA, Lingsma HF, Yoo AJ, Van Zwam WH, Van Oostenbrugge RJ, Van Der Lugt A, Dippel DWJ, et al. Associations of Ischemic Lesion Volume with Functional Outcome in Patients with Acute Ischemic Stroke: 24-Hour Versus 1-Week Imaging. Stroke. 2017;48:1233–1240. [DOI] [PubMed] [Google Scholar]
- 28.Federau C, Christensen S, Mlynash M, Tsai J, Kim S, Zaharchuk G, Inoue M, Straka M, Mishra NK, Kemp S, et al. Comparison of stroke volume evolution on diffusion-weighted imaging and fluid-attenuated inversion recovery following endovascular thrombectomy. Int. J. Stroke 2017;12:510–518. [DOI] [PubMed] [Google Scholar]
- 29.Olivot J-M, Mlynash M, Thijs VN, Kemp S, Lansberg MG, Wechsler L, Bammer R, Marks MP, Albers GW. Optimal Tmax threshold for predicting penumbral tissue in acute stroke. Stroke. 2009;40:469–475. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Olivot JM, Mlynash M, Inoue M, Marks MP, Wheeler HM, Kemp S, Straka M, Zaharchuk G, Bammer R, Lansberg MG, et al. Hypoperfusion intensity ratio predicts infarct progression and functional outcome in the DEFUSE 2 Cohort. Stroke. 2014;45:1018–1023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Core Team R. R: A Language and Enviroment for Statistical Computing. 2018; [Google Scholar]
- 32.Winzeck S, Hakim A, McKinley R, Pinto JAADSR, Alves V, Silva C, Pisov M, Krivov E, Belyaev M, Monteiro M, et al. ISLES 2016 and 2017-benchmarking ischemic stroke lesion outcome prediction based on multispectral MRI. Front. Neurol 2018;9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Lucas C, Kemmling A, Bouteldja N, Aulmann LF, Mamlouk AM, Heinrich MP. Learning to predict ischemic stroke growth on acute CT perfusion data by interpolating low -dimensional shape representations. Front. Neurol 2018;9:1–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Ho KC, Scalzo F, Sarma KV. El-Saden S, Arnold CW. A temporal deep learning approach for MR perfusion parameter estimation in stroke. Proc. - Int. Conf. Pattern Recognit 2016;0:1315–1320. [Google Scholar]
- 35.Kuang H, Qiu W, Boers AM, Brown S, Muir K, Majoie CBLM, Dippel DWJ, White P, Epstein J, Mitchell PJ, et al. Computed Tomography Perfusion – Based Machine Learning Model Better Predicts Follow-Up Infarction in Patients With Acute Ischemic Stroke. 2021;223–231. [DOI] [PubMed] [Google Scholar]
- 36.Demeestere J, Wouters A, Christensen S, Lemmens R, Lansberg MG. Review of Perfusion Imaging in Acute Ischemic Stroke. Stroke. 2020;51:1017–1024. [DOI] [PubMed] [Google Scholar]
- 37.Savitz SI, Baron JC, Yenari MA, Sanossian N, Fisher M. Reconsidering neuroprotection in the reperfusion era. Stroke. 2017;48:3413–3419. [DOI] [PubMed] [Google Scholar]
- 38.Rocha M, Jovin TG. Fast Versus Slow Progressors of Infarct Growth in Large Vessel Occlusion Stroke: Clinical and Research Implications. Stroke. 2017;48:2621–2627. [DOI] [PubMed] [Google Scholar]
- 39.Federau C, Mlynash M, Christensen S, Zaharchuk G, Cha B, Lansberg MG, Wintermark M, Albers GW. Evolution of volume and signal intensity on Fluid-attenuated inversion recovery MR images after endovascular stroke therapy. Radiology. 2016;280:184–192. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.