Table 2.
Model development using deep learning algorithms.
| References | Outcomes | Missing value | Major imaging pre-processing | Model architecture | Feature visualization | Validation |
|---|---|---|---|---|---|---|
| Hilbert et al. (34) | a. Good functional outcome (mRS ≤ 2) b. Successful reperfusion (TICI score≥2b) |
Patients with missing data were excluded | a. Brain extraction (50–400 HU) b. Rigid registration to a template c. Computing maximum intensity projection from 3D to 2D scans d. Normalization e. Imaging resampling (368 × 432) |
a. Functional outcome: supervised 2D-ResNet architecture with structured receptive field kernels model b. Successful reperfusion: a stacked denoising convolutional auto-encoder (2D-ResNet architecture with structured receptive field kernels) and fine-tuned model |
Gradient-weighted Class Activation Mapping | 4-fold cross validation |
| Samak et al. (35) | a. Good functional outcome (mRS ≤ 2) b. Individual mRS scores (0–6) |
Patients with missing data were excluded | a. Brain extraction (40–100 HU) b. Data augmentation (flip, rotations, elastic deformations, Gaussian noise) c. Normalization d. Imaging resampling (192x192x32) |
a. Multimodal model: image feature encoder, clinical metadata encoder, image and clinical metadata fusion b. 3D-convolutional kernels, attentional block |
n.a. | Hold-out validation |
| Nishi et al. (36) | Good functional outcome (mRS ≤ 2) | Patients with missing data were excluded | a. Brain extraction b. Data augmentation (rotations, translation, spatial scaling) c. Normalization d. ROIs labeling (ischemic core lesion) e. Imaging resampling (128 × 128 × 32) |
a. Multi-output model: A U-net segmentation task for imaging feature derivation, a 2-layer neural network for fine-tuning b. 3D-convolutional kernels |
Gradient-weighted Class Activation Mapping | 5-fold cross validation |
| Jiang et al. (37) | Hemorrhagic transformation (including HI1, HI2, PH1, and PH2) | Patients with missing data were excluded | a. Brain extraction b. Data augmentation (rotations, spatial scaling) c. ROIs labelling d. Imaging resampling (randomly cropped from ROIs) |
a. Multimodal model: multiple imaging feature encoders (DWI, MTT, and TTP), clinical metadata encoder, image and clinical metadata fusion b. 3D-based convolutional kernels, Inception V3 architecture |
n.a. | 5-fold cross validation |
mRS, Modified Rankin Scale; HI, hemorrhagic infarction; HU, Hounsfield Units; PH, parenchymatous hematoma; DWI, diffusion-weighted imaging; MTT, mean transit time; ROI: regions of interest; TTP, time to peak; TICI, thrombolysis in cerebral infarction score; n.a., not available.