PulmoSightCHEST X-RAY
01 / REVIEWNew review.
ImageNo image selected
Open a chest X-ray
Drop a chest X-ray here
PNG or JPG · up to 15 MBModel attention
Try a sampleChestDR examples
Loading…
MODEL PERFORMANCE
PENDINGAnalytics
How the model learned and performed.
50,000 additional images are verified and the expanded model is training. These charts and image results describe the completed model currently available.
About these charts
Loading measurements…
What the numbers mean
Training: images used to teach the model. Check set: separate images used to choose settings. Test set: images saved for the final evaluation.
Label agreement: how often each yes/no label matches the dataset. A high score can hide missed disease because most labels are negative.
Precision: how many displayed findings match the recorded labels. Recall: how many recorded findings the model catches. AUROC: a ranking score, where 0.5 is chance and 1 is best. None of these numbers is confidence in your X-ray.
The training charts show all saved rounds in order. Data and settings changed after rounds 5 and 8; the vertical markers show these changes. The later trial was not selected for the reader. Full records are available below.
The dataset labels can be wrong. Test results measure agreement with those labels, not a professional diagnosis.
More test details
Previous model / current model
Earlier ranking score—
Current ranking score—
AUROC measures ranking, not the percentage of correct diagnoses. Checkpoints are selected using validation data.
Training status
—Waiting for training status.
How reliable is “No Finding”?
This measures agreement with the CSV annotations, not clinical truth. “No Finding” must not be used to rule out disease. Its validation-selected threshold is separate from both disease thresholds and the top-score suggestion rule.
Results by finding
TEST SETROC AUC measures how well scores separate annotated positives from negatives. Sensitivity and specificity reflect the current reader, including its unverified fallback. AUROC and average precision measure the unchanged model scores. These numbers do not establish clinical safety.
Labels with fewer than 30 positive test images have limited evidence. Hernia, Pneumonia, and Fibrosis are excluded from the expanded model because each appears in fewer than 2% of training images.
No test metrics yet. Results appear only after a completed evaluation.
| Finding | Ranking (AUROC) | Precision score | Findings caught | Negatives correct | Test examples |
|---|
BEHIND THE READER
About PulmoSight.
A research tool for exploring chest X-rays.
PS
See the image.
Understand the limits.
PulmoSight scores 11 chest findings and “No Finding,” with overlays showing what influenced the model.
It can miss disease or flag something that is not there. It is not a diagnosis.
Reading your results
More than one finding can appear. A low score does not rule out disease. “No Finding” is not an all-clear. When no screening threshold passes, the reader shows a top-ranked, unverified suggestion.
Overlays show influential regions, not confirmed lesions. Hernia, Pneumonia, and Fibrosis are not evaluated.
Model & training
The active model was trained on 100,000 NIH chest X-rays. It uses InceptionV3 with 11 disease outputs and a separate No Finding output.
Download model record ↗Checking model…
Privacy & supported images
Your X-ray is processed in this browser. Uploaded images are not saved or sent to an inference server.
Use de-identified, frontal PNG or JPG chest X-rays. DICOM files, lateral views, pediatric use, and photographs of films are outside the evaluated scope.
Windows and Android downloads open PulmoSight in an app window. iPhone and iPad use Safari’s Add to Home Screen option. An internet connection is needed to open the app and load the model.
Data & sources
NIH labels come from reports and can contain errors. ChestDR provides radiologist-reviewed labels for nine matching targets; Mass, Infiltration and No Finding are unannotated there.
Analytics includes the full training history, held-out comparisons and the latest experimental model’s release decision.