RSNA-ICH partial re-derivation / non-commercial
bv-headct-v1 · BlackVoxel
Intracranial hemorrhage on CT.
Multi-label classification of intracranial hemorrhage and five subtypes.
What it does
The model analyzes one axial slice and produces six outputs: intracranial hemorrhage and five subtypes. The demo shows the attention map and draft.
- Model
- bv-headct-v1
- Ownership
- BlackVoxel
- Base
- ResNet-18 / ImageNet transfer / 6 labels
- Data
- RSNA-ICH partial re-derivation / non-commercial
Demo
Image, attention map and draft report.
The screen uses previously processed results. Inference is not run on this page.
Intraparenchymal hemorrhage · 100%
AXIALW/L · AUTO
STORED MODEL OUTPUTGRAD-CAM
Training and evaluation
Data and setup.
ResNet-18 / ImageNet transfer / 6 labels
Task-specific labels
six labels
ResNet-18 fine-tuned for six independent outputs using a partial, non-random RSNA-ICH subset. The checkpoint was selected on validation and evaluated on the test set.
Macro-AUROC 0.932 across six labels. Sensitivity 0.827 for any hemorrhage on the 4,467-slice test set.
Test results
Measured performance.
macro-AUROC
six labelssensitivity
any hemorrhagetest slices
partial sampleThe result comes from a partial, non-random sample. The dataset has a non-commercial license.
Limitations
Scope of this result.
- 01
Non-commercial data and a partial non-random sample.
- 02
It analyzes one 2D slice, not the complete volume.
- 03
Visual attention does not localize or measure the hemorrhage.
Next model
CT / CHEST