BlackVoxel models
Models for
medical imaging.
BlackVoxel develops proprietary models for specific imaging tasks. Each page lists the data, training method, test result and current limitations.
Available models
The metrics below come from the test sets named on each page. They are research results, not clinical validation.
DX / CXR 01 Chest X-ray
Chest finding classification with an attention map and report draft.
DX / MSK 02 Limb fracture
Fracture classification on limb radiographs with Grad-CAM attention.
MR / BRAIN 03 Brain MRI tumor
MRI-slice classification across no tumor, glioma, meningioma and pituitary tumor.
US / BREAST 04 Breast ultrasound
Classification across normal, benign nodule and malignant nodule on BUSI.
CT / HEAD 05 Head CT hemorrhage
Multi-label classification of intracranial hemorrhage and five subtypes.
CT / CHEST 06 Lung nodule characterization
Binary characterization of a crop centered on a lung nodule.
US / ABD 07 Abdominal organ recognition
Classification of ten structures in abdominal ultrasound frames.
MG / ROI 08 Mammography ROI characterization
Benign or malignant classification on CBIS-DDSM finding crops.
MR / SPINE 09 Lumbar MRI findings
Multi-label classification on one midsagittal lumbar MRI slice.
Method
How each model is trained and tested.
- 01Task
We define the input, labels and exclusions.
- 02Training
We train the model or fine-tune a pretrained base.
- 03Evaluation
The checkpoint is selected on validation. Reported metrics come from the test set.
- 04Review
Results are shown in the viewer for physician review.
AUROC reports performance on the test set.
The models are proprietary to BlackVoxel. Public datasets and pretrained bases remain under their original licenses. The metrics are research results and do not represent regulatory clearance.