PneumoShift: kids_plus_adults

Research model. Not a medical device. Not for diagnosis.

A resnet18.tv_in1k image classifier with one output: the chance that a frontal chest X-ray shows pneumonia. Part of a student project that asks whether a model trained on children still works on adults.

Training data

• 4578 children X-rays (Kermany train)
• 6000 adult X-rays (RSNA adapt)

Children: Kermany et al. 2018 (ages 1 to 5, one hospital in Guangzhou, China, CC BY 4.0). Adults, if listed above: RSNA Pneumonia Detection Challenge 2018 (USA).

Intended use

Flag chest X-rays for priority review by a clinician, in a research or teaching setting. The model must not be used to diagnose or treat anyone.

How to use

from pneumoshift.inference import Predictor
r = Predictor("RL8596/pneumoshift-kids-plus-adults").predict(pil_image)   # probability, flagged, heatmap, ood_level, ...

• Probability = sigmoid(logit / 1.043) (temperature scaling).
• An X-ray is flagged when the probability is at least 0.1575 (95% sensitivity on the validation set).
• ood.npz, if present, holds the statistics for the "unlike my training data" warning.

Test results

Limitations

• It only learned from the patient groups and hospitals listed under training data. For other age groups, hospitals or X-ray machines it can be confidently wrong.
• It knows two answers, pneumonia or not, and does not recognise other findings.
• The out-of-distribution warning catches many unusual X-rays, not all of them.
• Not checked by any authority.