Pretrained models for automatic DWI-FLAIR mismatch classification in acute ischemic stroke.
Trained on the SOOP dataset (OpenNeuro ds004889) using a 3D DenseNet121 architecture (MONAI).
Labels generated via the rDWI-IQR method (Scheldeman et al., 2022).
Preprocessing: TRSAA registration, 2×2×2mm isotropic resampling, Z-score normalization, CropOrPad 80×112×96.
Config: AdamW · lr=1e-3 · wd=1e-5 · warmup cosine · online augmentation
Val AUC=0.849 | Test AUC=0.834 | External AUC (ISLES 2022)=0.813
Preprocessing: resize 256×256×36, min-max normalization, TRSAA registration (Kim et al., 2025).
Config: AdamW · lr=1e-3 · wd=1e-5 · warmup cosine · online augmentation
Val AUC=0.790 | Test AUC=0.778 | External AUC (ISLES 2022)=0.802
1import torch
2from huggingface_hub import hf_hub_download
3from monai.networks.nets import DenseNet121
4
5model = DenseNet121(spatial_dims=3, in_channels=2, out_channels=2, dropout_prob=0.5)
6
7path = hf_hub_download(repo_id="your_username/dwi-flair-mismatch", filename="best_own_model.pth")
8model.load_state_dict(torch.load(path, map_location="cpu"))
9model.eval()
