FracAtlas / CC BY 4.0
bv-limbfx-v1 · BlackVoxel
Fracture classification on radiographs.
Fracture classification on limb radiographs with Grad-CAM attention.
What it does
The model estimates fracture probability on a limb radiograph. Grad-CAM shows the regions that most influenced the output.
- Model
- bv-limbfx-v1
- Ownership
- BlackVoxel
- Base
- ResNet-18 / ImageNet transfer
- Data
- FracAtlas / CC BY 4.0
Demo
Image, attention map and draft report.
The screen uses previously processed results. Inference is not run on this page.
Fracture · 99.8%
DXW/L · AUTO
STORED MODEL OUTPUTGRAD-CAM
Training and evaluation
Data and setup.
ResNet-18 / ImageNet transfer
Task-specific labels
FracAtlas test
ResNet-18 initialized with ImageNet weights and fine-tuned on FracAtlas. The checkpoint was selected on validation and measured once on the test set.
Held-out test with 613 radiographs, including 108 fracture cases. Classifier AUROC 0.887.
Test results
Measured performance.
AUROC
FracAtlas testsensitivity
operating pointspecificity
same thresholdAt the evaluated threshold, sensitivity was 0.67 and specificity was 0.96.
Limitations
Scope of this result.
- 01
Result from a public benchmark without external Brazilian validation.
- 02
The box follows Grad-CAM attention and is not a fracture segmentation.
- 03
A single view does not replace study review and clinical context.
Next model
MR / BRAIN