CBIS-DDSM / digitized film / CC BY 4.0
bv-mammo-v1 · BlackVoxel
Mammography ROI characterization.
Benign or malignant classification on CBIS-DDSM finding crops.
What it does
The model classifies a region-of-interest crop as benign or malignant. The region is provided in advance.
- Model
- bv-mammo-v1
- Ownership
- BlackVoxel
- Base
- ResNet-18 / ImageNet transfer / ROI crop
- Data
- CBIS-DDSM / digitized film / CC BY 4.0
Demo
Image, attention map and draft report.
The screen uses previously processed results. Inference is not run on this page.
Suspicious lesion · 98.9%
CCW/L · AUTO
STORED MODEL OUTPUTGRAD-CAM
Training and evaluation
Data and setup.
ResNet-18 / ImageNet transfer / ROI crop
Task-specific labels
patient-grouped test
ResNet-18 fine-tuned on 2,437 finding crops. Patient-grouped splitting keeps cases from the same patient out of different partitions.
AUROC 0.748, AUPRC 0.701 and sensitivity 0.626 on the 627-crop test set.
Test results
Measured performance.
AUROC
patient-grouped testAUPRC
malignant classsensitivity
argmax thresholdThe AUROC of 0.748 refers to characterization of the provided ROI. The evaluation does not include detection or screening.
Limitations
Scope of this result.
- 01
Digitized film, not full-field digital mammography.
- 02
The region is already provided; the model does not detect the lesion.
- 03
Sensitivity 0.626 leaves a relevant share of positives unflagged.
Next model
MR / SPINE