SPIDER / CC BY 4.0
bv-lumbarmr-v1 · BlackVoxel
Degenerative findings on lumbar MRI.
Multi-label classification on one midsagittal lumbar MRI slice.
What it does
The model produces six outputs from one midsagittal slice. Results are study-level, without localization by disc.
- Model
- bv-lumbarmr-v1
- Ownership
- BlackVoxel
- Base
- ResNet-18 / ImageNet transfer / 6 labels
- Data
- SPIDER / CC BY 4.0
Demo
Image, attention map and draft report.
The screen uses previously processed results. Inference is not run on this page.
Disc bulging · 92.6%
MID-SAGW/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 on one midsagittal slice per patient. Per-disc labels were rolled up into six study-level outputs.
Macro-AUROC 0.762 on the 32-patient test set. Disc bulging AUROC was 0.931; herniation sensitivity was 0.222.
Test results
Measured performance.
macro-AUROC
six labelsbulging AUROC
common labelherniation sens.
rare labelThe test set is small and rare labels are weak. The model cannot rule out herniation or localize a disc.
Limitations
Scope of this result.
- 01
One midsagittal slice, not the full sagittal and axial study.
- 02
Study-level output without L1-L2 to L5-S1 localization.
- 03
Test with 32 patients and low sensitivity for rare labels.
Next model
DX / CXR