NIH ChestX-ray14
bv-chesxr-v1 · BlackVoxel
Chest findings in MIMPS.
Chest finding classification with an attention map and report draft.
What it does
bv-chesxr-v1 analyzes chest radiographs in MIMPS. The viewer displays findings, the attention map and a draft for physician review.
- Model
- bv-chesxr-v1
- Ownership
- BlackVoxel
- Base
- DenseNet-121 / torchxrayvision
- Data
- NIH ChestX-ray14
Demo
Image, attention map and draft report.
The screen uses previously processed results. Inference is not run on this page.
Pneumonia · 75.3%
APW/L · AUTO
STORED MODEL OUTPUTGRAD-CAM
Training and evaluation
Data and setup.
DenseNet-121 / torchxrayvision
Task-specific labels
NIH test
bv-chesxr-v1 uses a pretrained chest radiography base. BlackVoxel developed the task adaptation, evaluation and MIMPS integration. Public components remain under their original licenses.
On the NIH ChestX-ray14 test set, macro-AUROC was 0.8171 across 14 chest findings.
Test results
Measured performance.
macro-AUROC
NIH testbenchmark labels
multi-label outputstatus
non-diagnosticMacro-AUROC measures discrimination on NIH ChestX-ray14. The evaluation does not include calibration or data from a Brazilian institution.
Limitations
Scope of this result.
- 01
Public pretrained base, retained under its original license.
- 02
The score is not calibrated and Grad-CAM does not delineate a lesion.
- 03
Every output requires full-study reading and physician review.
Next model
DX / MSK