Multi-label chest pathology assessment, pneumothorax localization and cardiothoracic ratio
Research tool — not a regulated medical device. Their results are not a diagnosis and can be wrong. Every result must be reviewed by a qualified clinician, who remains responsible for the clinical decision.
Pathology scoring: TorchXRayVision (multi-label chest radiograph classification). License: Apache 2.0.
Cardiothoracic ratio: a separate segmentation model outlines both lungs and the heart, then measures the cardiac and thoracic widths.
Pneumothorax localization uses a separate segmentation model on this deployment.
Running the AI now returns, in addition to the pathology scores and the pneumothorax mask:
Screens a chest X-ray for common findings (consolidation, effusion, pneumothorax, cardiomegaly and more) with a multi-label CNN, returning per-finding probabilities and a heat-map of the regions driving each call.
AI result
Pneumothorax highlighted in cyan
Lungs and heart segmented — CTR 0.48, within normal limits
| Pneumothorax | 92% | |
| Infiltration | 11% | |
| Effusion | 6% | |
| Fracture | 5% |
CTR
0.48
Within normal limits
Lung fields R / L
54% / 46%
Asymmetry: 7.1%
Pneumothorax extent
~28%
of the right hemithorax
Widths and areas in centimetres are added when the study carries a pixel spacing — this example is a PNG, so only ratios can be measured.