Table 1.
Study characteristics, model methodology, and predictive performance of the included studies.
| First author and year | Study characteristics | ||||||
|---|---|---|---|---|---|---|---|
| Dataset source |
Inclusion/
exclusion criteria |
Total patient number (n) | Training and validation sets (n) | Test set (n) | External test set | Multivendor images | |
| Gottrup et al. (24) | Single center | N | 14 | Leave-one-out cross validation | N | NR | |
| McKinley et al. (25) | Single center | Y | 61 | 25 | 36 | N | N |
| Livne et al. (26) | Multicenter (I-KNOW study and the Ischemic Preconditioning trial) | Y | 195 | ≈156 | ≈39 | N | Y |
| Nielsen et al. (14) | Multicenter (I-KNOW and remote ischemic preconditioning studies) | Y | 222 | 187 | 35 | N | Y |
| Pinto et al. (27) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Winzeck et al. (28) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Clèrigues et al. (29) | Multicenter (ISLES 2018 dataset) | Y | 103 | 63 | 40 | N | Y |
| Ho et al. (30) | Single center | Y | 48 | ≈43 | ≈5 | N | N |
| Kasasbeh et al. (31) | Multicenter | Y | 103 | ≈82 | ≈21 | N | NR |
| Pérez Malla et al. (32) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Robben et al. (33) | Multicenter (MR CLEAN study) | Y | 188 | ≈150 | ≈38 | N | NR |
| Winder et al. (34) | Single center | Y | 90 | Leave-one-out cross validation | N | N | |
| Grosser et al. (35) | Multicenter | Y | 99 | Leave-one-out cross validation | N | NR | |
| Grosser et al. (36) | Multicenter | Y | 99 | Leave-one-out cross validation | N | NR | |
| Hu et al. (37) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Kim et al. (38) | Single center | Y | 92 unsuccessful recanalization 36 and successful recanalization 56 | 53 | 39 | N | N |
| Kumar et al. (39) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Pinto et al. (40) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Qiu et al. (41) | Single center | Y | 257 | 157 | 100 | N | N |
| Wang et al. (42) | Multicenter (ISLES 2018 dataset) | Y | 103 | 63 | 40 | N | Y |
| Yu et al. (17) | Multicenter (ICAS and DEFUSE-2 studies) | Y | 182 | ≈146 | ≈36 | N | NR |
| Benzakoun et al. (11) | Single center | Y | 394 | ≈358 | ≈36 | N | N |
| Debs et al. (43) | Multicenter (HIBISCUS-STROKE and I-KNOW cohorts) | Y | 109 reperfused 74 and non-reperfused 35 | Reperfused≈69 and non-reperfused≈28 | Reperfused≈15 and non- reperfused≈ 7 | N | NR |
| Hakim et al. (44) | Multicenter (ISLES 2018 dataset) | Y | 103 | 63 | 40 | N | Y |
| Hokkinen et al. (45) | Single center | Y | 83 | NR | NR | N | N |
| Hokkinen et al. (46) | Single center | Y | 89 | None | 89 | N | N |
| Klug et al. (47) | Single center | Y | 144 intravenous thrombolysis (IVT) 80, endovascular thrombectomy (EVT) 64 | ≈115 | ≈29 | N | N |
| Kuang et al. (13) | Multicenter (Prove-IT study and HERMES collaboration) | Y | 205 | 68 | 137 | Y | NR |
| Modrau et al. (48) | Multicenter (TEA-Stroke Trial) | Y | 52 theophylline 27 and control group 25 | NR | NR | N | NR |
| Pinto et al. (49) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Qiu et al. (15) | Multicenter (Prove-IT study) | Y | 196 | 170 | 26 | N | NR |
| Soltanpour et al. (50) | Multicenter (ISLES 2018 dataset) | Y | 103 | 63 | 40 | N | Y |
| Vupputuri et al. (51) | Multicenter (ISLES 2017 dataset) | Y | 75 | 43 | 32 | N | N |
| Yu et al. (16) | Multicenter (ICAS, DEFUSE and DEFUSE-2 studies) | Y | 185 | 118 | 67 | N | NR |
| He et al. (12) | Single center | Y | 70 | 59 | 11 | N | N |
| Lin et al. (52) | Single center | Y | 261 | ≈209 | ≈52 | N | NR |
| Shi et al. (53) | Multicenter (ISLES 2018 dataset) | Y | 103 | 63 | 40 | N | Y |
| Zhu et al. (54) | Multicenter | N | 89 | ≈71 | ≈18 | N | N |
| Model methodology | Predictive performance | ||||||
| First author and year | Summary of the model | Input parameters | Ground Truth |
Primary metric (DSC score) |
Secondary metrics | ||
| Gottrup et al. (24) | k-nearest neighbor classification | MR-CBF, CBV, MTT, DWI, ADC, T2WI | Infarct lesions manually segmented on follow-up T2WI 5 days or later | NR | AUC: 0.814 ± 0.001 Sensitivity: 0.73 Specificity: 0.73 |
||
| McKinley et al. (25) | Random forest classifier, including segmentation and predictive classifiers | Features extracted from MR-T1 contrast, T2WI, ADC, CBF, CBV, TTP, Tmax | Final infarct lesions manually segmented on follow-up T2WI at 90 days by 2 radiologists | 0.34 ± 0.22 | AUC: 0.94 ± 0.08 Sensitivity: 0.52 Specificity: 0.99 Precision: 0.56 |
||
| Livne et al. (26) | Extreme gradient boosting (XGBoost) | MR-DWI, T2-FLAIR, and TTP derived from the concentration curve; CBF, MTT and Tmax using oscillatory singular value decomposition deconvolution; CBF, CBV, MTT, Tmax, relative transit time heterogeneity and capillary transit time heterogeneity using a statistical approach | Final infarct lesions semi-automatically segmented on follow-up T2-FLAIR | NR | AUC: 0.92 Accuracy: 0.84 |
||
| Nielsen et al. (14) | Modified SegNet | MR-mean capillary transit time, CBV, CBF, cerebral metabolism of oxygen, relative transit time heterogeneity, delay, TRACE DWI, ADC, and T2-FLAIR | Infarcts lesions manually segmented on follow-up T2-FLAIR at 30 days by 4 expert radiologists | NR | AUC: 0.88 ± 0.12 | ||
| Pinto et al. (27) | Fully convolutional U-Net combined with a 2D-dimensional gated recurrent unit layer | MR-ADC, rCBF, rCBV, MTT, TTP, Tmax and clinical information-TICI score | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.29 ± 0.22 | Precision: 0.26 ± 0.23 Recall: 0.61 ± 0.28 |
||
| Winzeck et al. (28) | Multiscale U-net architecture trained with negative Dice score | MR-ADC, rCBF, rCBV, MTT, TMAX, TTP, Raw PWI and clinical information time-since-stroke, time-to-treatment, TICI and mRS scores | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.31 ± 0.23 | Sensitivity: 0.45 ± 0.31 Precision: 0.36 ± 0.27 |
||
| Clèrigues et al. (29) | 2D asymmetrical residual encoder–decoder CNN by using a more regularized network training procedure, symmetric modality augmentation and uncertainty filtering | CT-raw CTP series and CBF, CBV, MTT, Tmax | Infarct core manually segmented by a single investigator and then subjected to group review until acceptance | 0.547 ± 0.242 | Sensitivity: 0.609 ± 0.250 | ||
| Ho et al. (30) | Unit CNN-contralateral model including modified input patches (patches of interest paired with contralateral patches), convolutional layer architecture and unit temporal filter learning | MR-PWI source image | Infarct lesions semi-automatically segmented on follow-up FLAIR at 3–7 days by a radiologist | NR | AUC: 0.871 ± 0.024 Precision: 0.222 Recall: 0.799 |
||
| Kasasbeh et al. (31) | Feed-forward ANN | CT-rCBF, CBV, MTT, and Tmax | acute infarct lesions segmented on follow-up DWI at median time delay of 40.5 min | 0.48 (IQR 0.23–0.70) | AUC: 0.85 Mean volume error: 13.8 ± 13.6 ml |
||
| Pérez Malla et al. (32) | DeepMedic model with PReLU activation using transfer learning, data augmentation and binary morphological post-processing operations | MR-ADC, MTT, and rCBF | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.34 | - | ||
| Robben et al. (33) | Fully convolutional network with PReLU activation | CT-native CTP, downsampled CTP, arterial input function and clinical data-time between stroke onset and imaging, time between imaging and the end of the mechanical thrombectomy, mTICI score and persistence of occlusion at 24 h | Infarct lesions semi-automatically segmented on follow-up NCCT at 1–5 days by an experienced reader | 0.48 | Mean absolute volume error: 36.7 ml | ||
| Winder et al. (34) | Random forest classifier | MR-ADC, distance to ischemic core, tissue type, anatomical location, CBV, MTT, Tmax, CBF and clinical data-NIHSS, age, sex, and time from symptom onset | Final infarct lesion manually segmented on FLAIR or DWI or NCCT at 5–7 days by an experienced medical expert | 0.447 ± 0.247 | - | ||
| Grosser et al. (35) | Random forest classifier trained by local and global approaches | MR-ADC, CBF, CBV, MTT, Tmax | Infarct lesions manually segmented on follow-up FLAIR at 1–7 days by 2 neurologists in consensus | 0.353 ± 0.220 | AUC: 0.859 ± 0.089 Sensitivity: 0.415 ± 0.231 Specificity: 0.964 ± 0.034 | ||
| Grosser et al. (36) | XGBoost | MR-ADC, CBF, CBV, MTT, Tmax and voxel-wise lesion probabilities | Infarct lesions manually segmented on follow-up FLAIR within 7 days by 2 neuroradiologists in consensus | 0.395 ± 0.229 | AUC: 0.888 ± 0.101 | ||
| Hu et al. (37) | Brain SegNet: a 3D dense segmentation network based on ResNet and trained with data augmentation and Focal loss | MR-TTP, Tmax, rCBV, rCBF, MTT, ADC | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.30 ± 0.22 | Precision: 0.35 ± 0.27 Recall: 0.43 ± 0.27 | ||
| Kim et al. (38) | Random forest classifier | Features derived from MR-ADC and rTTP: range, mean, median, min, max, standard deviation, skew, kurtosis, 10 th percentile, 25 th percentile, 75 th percentile, and 90 th percentile | Infarct lesions manually segmented on follow-up DWI at 7 days | 0.49 (IQR 0.37–0.59) | Unsuccessful recanalization: AUC: 0.746 ± 0.048 Mean volume error: −32.5 ml Successful recanalization: AUC: 0.764 ± 0.127 Mean volume error: 3.5 ml | ||
| Kumar et al. (39) | Classifier-Segmenter network, using a hybrid training strategy with a self-similar (fractal) U-Net model | MR-DWI, ADC, CBV, CBF, MTT, TTP, Tmax | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.28 ± 0.22 | Precision: 0.37 ± 0.29 Recall: 0.45 ± 0.34 | ||
| Pinto et al. (40) | Two-branch Restricted Boltzmann Machine provides lesion and hemodynamics features from parametric MRI maps, then combined with parametric MRI maps and fed to a U-net using NReLU activation | MR-ADC, MTT, TTP, rCBF and rCBV | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.38 ± 0.22 | Precision: 0.41 ± 0.26 Recall: 0.53 ± 0.29 | ||
| Qiu et al. (41) | Random forest classifier | Features derived from NCCT: Hounsfield units, bilateral density difference, hypoattenuation measurement, distance feature, atlas-encoded lesion location feature | Early infarct lesions manually segmented on follow-up DWI within 1 h | NR | Mean volume error: 11 ml | ||
| Wang et al. (42) | CNN model with a feature extractor, a pseudo-DWI generator and a final lesion segmenter using hybrid loss function | CT-CBF, CBV, MTT, Tmax and synthesized pseudo-DWI | Infarct core manually segmented by a single investigator and then subjected to group review until acceptance | 0.54 ± 0.21 | Precision: 51.20 ± 22.00 Recall: 64.20 ± 23.99 | ||
| Yu et al. (17) | 2.5D attention-gated U-Net using mixed loss functions | MR-DWI, ADC, Tmax, MTT, CBF, CBV | Final infarct lesions manually segmented on follow-up T2-FLAIR at 3–7 days by a neuroradiologist | 0.53 (IQR 0.31–0.68) | AUC: 0.92 (IQR 0.87–0.96) Mean volume error: 9 ml (IQR −14ml−29ml) | ||
| Benzakoun et al. (11) | Gradient Boosting | MR-DWI, ADC, Tmax, MRR, CBF, CBV | Infarct lesions manually segmented on follow-up DWI around 24 h by a neuroradiologist | 0.53 (IQR 0.29–0.68) | AUC: 0.98 (IQR 0.95–0.99) Mean volume error: 27.7 ± 40.3 ml |
||
| Debs et al. (43) | U-Net with multi-class Dice loss functions | MR-DWI, ADC, Tmax, CBF, CBV | Final infarct lesions semi-automatically segmented on follow-up T2-FLAIR at 6- or 30-day using intensity-based thresholding method | Reperfused: 0.44 ± 0.25 Non-reperfused: 0.47 ± 0.17 |
Reperfused: AUC: 0.87 ± 0.13 Precision:0.50 ± 0.27 Recall:0.50 ± 0.26 Non-reperfused: AUC: 0.81 ± 0.13 Precision: 0.49 ± 0.22 Recall: 0.52 ± 0.21 |
||
| Hakim et al. (44) | 3D multi-scale U-shape network with atrous convolution | CT-CTP source data, CBF, CBV, MTT, Tmax | Infarct core manually segmented by a single investigator and then subjected to group review until acceptance | 0.51 ± 0.31 | Mean absolute volume error: 10.24 ± 9.94 ml Precision: 0.55 ± 0.36 Recall: 0.55 ± 0.34 |
||
| Hokkinen et al. (45) | 3D CNN | CT-CTA source image | Infarct lesions manually segmented on follow-up CT with median time interval of 36 h | NR | Mean volume error: −16.3 ml | ||
| Hokkinen et al. (46) | 3D CNN | CT-CTA source image | Infarct lesions manually segmented on follow-up CT or DWI within 5 days by a radiologist | NR | Mean volume error: 13.9 ± 12.5 ml | ||
| Klug et al. (47) | General linear regression model | CT-MTT, Tmax, CBF and CBV and multi-perfusion parameter analysis | Final infarct lesions segmented on T2-FLAIR within 10 days by 2 neuroradiologists | 0.155 | AUC: 0.89 Volume error: IVT: 4.6 ml (IQR 0.7–19.9), EVT: 32.8 ml (IQR 8.9–64.7) |
||
| Kuang et al. (13) | Random forest classifier | CT-average map, Tmax, CBF, CBV and clinical data-onset-to-imaging time, imaging-to-reperfusion time | PRoveIT study: infarct lesions manually segmented on follow-up DWI or NCCT by 2 experts in consensus; HERMES collaboration: infarct lesions automatically segmented followed by manual corrections | 0.388 (IQR 0.192–0.541) | AUC: 0.81 ± 0.11 Volume error: −3.2 ml (IQR −16.7–6.1) |
||
| Modrau et al. (48) | Random forest classifier | MR-ADC, CBF, CBV, MTT, Tmax, tissue type probability, anatomical location, distance to the ischemic core and clinical data-age, sex, baseline NIHSS, time of stroke onset to medical application | Infarct lesions manually segmented on follow-up T2-FLAIR at 24 h | Theophylline subgroup: 0.40 ± 0.249 Placebo subgroup: 0.35 ± 0.243 |
|||
| Pinto et al. (49) | 2D U-Net with a data-driven branch computing spatio-temporal features from DSC-MRI | MR-DSC-MRI spatio-temporal information, Tmax, TTP, MTT, rCBV, rCBF, ADC | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.31 ± 0.21 | Precision: 0.29 ± 0.23 Recall: 0.63 ± 0.30 |
||
| Qiu et al. (15) | Random forest classifier | Features derived multi-phase CTA: average and standard deviation of HUs across 3-phase CTA images, coefficient of variance of HUs in 3-phase CTA images, changing slopes of HUs between any two phases, peak of HUs in 3-phase CTA images, time of peak HU | Infarct lesions manually segmented on follow-up DWI/NCCT at 24/36h by 2 radiologists | 0.247 (IQR 0.138–0.304) | Mean volume error: 21.7 ml | ||
| Soltanpour et al. (50) | MultiRes U-Net | CT-CBF, CBV, MTT, Tmax, contrast map, Tmax heatmap | Infarct core manually segmented by a single investigator and then subjected to group review until acceptance | 0.68 ± 0.26 | Sensitivity: 0.68 ± 0.15 Mean absolute volume error: 22.62 ± 7.3 ml | ||
| Vupputuri et al. (51) | MCN-DN: Multi-path convolution leveraged attention deep network with LReLU | MR-ADC, CBF, CBV, MTT, TTP | Final infarct lesions manually segmented on follow-up T2WI at 90 days by a neuroradiologist | 0.47 | Sensitivity:0.867 Specificity:0.972 |
||
| Yu et al. (16) | Attention-gated U-Net with mixed loss functions | MR-DWI, ADC, Tmax, MTT, CBV, CBF and masks of Tmax (>6s) and ADC (620 × 10-6 mm2/s ) | iCAS and DEFUSE-2 studies: final infarct lesions segmented on T2-FLAIR at 3–7 days; DEFUSE study: final infarct lesions segmented on T2-FLAIR at 30 days | 0.57 (IQR 0.30–0.69) | AUC: 0.94 (IQR 0.89–0.97) Volume error: 0 ml (IQR -44–81) |
||
| He et al. (12) | 2D U-Net with binary focal loss and Jaccard loss combined functions | CT-CBF, CBV, MTT, Tmax | Infarct lesions manually segmented on follow-up DWI/SWI or NCCT | 0.61 | AUC: 0.92 Sensitivity: 0.63 Specificity: 0.98 Absolute volume error: 26.5 ml (IQR 9.9–31.7) |
||
| Lin et al. (52) | R2U-RNet with residual refinement unit (RRU) activation and multiscale focal loss functions | CT-NCCT with intensity normalization and histogram equalization | Infarct lesion manually segmented on follow-up DWI within 7 days by a radiologist | 0.54 ± 0.29 | - | ||
| Shi et al. (53) | C2MA-Net: a cross-modal cross-attention network | CT-CBF, CBV, MTT, Tmax | Infarct core manually segmented by a single investigator and then subjected to group review until acceptance | 0.48 | Precision: 0.48 Recall: 0.59 |
||
| Zhu et al. (54) | ISP-Net: a multi-scale atrous convolution with weighted cross entropy loss functions | CT-CTP source data, CBF, CBV, MTT, Tmax | Infarct lesions segmented on follow-up CT or DWI at 1-7 days | 0.801 ± 0.078 | AUC: 0.721 ± 0.108 Specificity:0.995 ± 0.002 Precision: 0.813 ± 0.066 Recall: 0.795 ± 0.115 |
||
Y, Yes; N, No; NR, not reported; PReLU, parametric rectified linear unit; NReLU, noisy rectified linear unit; LReLU, leaky rectified linear unit; CTA, CT angiography; CTP, CT perfusion; DWI, diffusion-weighted imaging; ADC, apparent diffusion coefficient; CBF, cerebral blood flow; CBV, cerebral blood volume; MTT, mean transit time; TTP, time to peak; Tmax, time to maximum of the residue function; NIHSS, National Institute of Health stroke scale; mTICI, modified thrombolysis in cerebral infarction; mRS, modified Rankin scale; DSC, dice similarity coefficient; AUC, area under the receiving operator characteristic curve.
Studies included in the meta-analysis were presented in bold font.