Variable Generalization Performance of a Deep Learning Model to Detect Pneumonia in Chest Radiographs¶
Zech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & Oermann, E. K. (2018). Variable Generalization Performance of a Deep Learning Model to Detect Pneumonia in Chest Radiographs: A Cross-Sectional Study. PLOS Medicine, 15(11).
Cited by¶
4 citations across 4 artifacts.
Each citation links to the sentence it supports in the citing article.
Primes¶
- Cue Outcome Decoupling
- Make it concrete: a pneumonia detector trained at one hospital latches onto a scanner watermark — a token burned into images by that site's portable X-ray machine, which was used disproportionately on already-sick inpatients, so the watermark correlated with disease in training.
This sourceDocuments a pneumonia detector keying on a scanner/hospital signal correlated with disease prevalence, failing across sites.
Supported in partVerified against the work's full text
Zech et al. show CNNs picking up scanner and department cues tied to pneumonia prevalence, but only partly the specific watermark-on-sick-inpatients story told here.
“CNNs could separate portable radiographs from the inpatient wards and emergency department in MSH data with 100% accuracy and that these patient groups had significantly different prevalences of pneumonia”
- Make it concrete: a pneumonia detector trained at one hospital latches onto a scanner watermark — a token burned into images by that site's portable X-ray machine, which was used disproportionately on already-sick inpatients, so the watermark correlated with disease in training.
- Out Of Distribution Detection
- A single worked instance shows the move: a chest-X-ray classifier trained on adults, deployed at a paediatric clinic, confidently mis-diagnoses paediatric anatomy until a retrofit OOD detector flags those inputs as out-of-distribution and routes them to a human radiologist — in-distribution performance unchanged, the silent failure converted into an explicit referral.
This sourceShows chest-radiograph CNNs exploit site-specific confounders and generalize poorly to new hospitals/populations, so on-site performance overstates real-world reliability — empirical grounding for distribution shift degrading clinical imaging models and motivating caution/human review on shifted inputs.
- A single worked instance shows the move: a chest-X-ray classifier trained on adults, deployed at a paediatric clinic, confidently mis-diagnoses paediatric anatomy until a retrofit OOD detector flags those inputs as out-of-distribution and routes them to a human radiologist — in-distribution performance unchanged, the silent failure converted into an explicit referral.
- Relevance Substitution
- During training the model discovers a substituted channel: a metal token or scanner artefact that hospitals with sicker populations happen to stamp on their films correlates with the label, so the model's inference runs through the artefact.
This sourceShows a pneumonia classifier keying on scanner/hospital artefacts rather than lung pathology, collapsing on out-of-distribution equipment.
SupportedVerified against the work's full text
Zech et al. show a CNN detecting a site-specific metal token, link site and scanner features to disease prevalence, and report that models exploit them to predict pneumonia indirectly.
“the CNN has learned to detect a metal token that radiology technicians place on the patient in the corner of the image field of view at the time they capture the image. When these strong features are correlated with disease prevalence, models can leverage them to indirectly predict disease.”
- During training the model discovers a substituted channel: a metal token or scanner artefact that hospitals with sicker populations happen to stamp on their films correlates with the label, so the model's inference runs through the artefact.
- Shortcut Learning
- The identical structure governs medical imaging: a pneumonia detector trained across hospitals latches onto the X-ray machine's metadata or image-edge tokens that correlate with the sicker hospital's higher base rate, rather than the lung pathology; it collapses at a new hospital, and the fix is multi-site training that varies the scanner independently of the diagnosis.
This sourceShows a pneumonia detector learning hospital/scanner-specific artifacts that correlate with disease prevalence, collapsing at a new site — the medical-imaging shortcut.
Supported in partVerified against the work's full text
Zech et al. show CNNs exploit hospital/scanner cues tied to prevalence and generalize unevenly, but their pooled multi-site training did not fix this, and degradation was partial, not a collapse.
“By engineering cohorts of varying prevalence, we demonstrated that the more predictive a hospital system was of pneumonia, the more it was exploited to make predictions, which led to poor generalization on external datasets.”
- The identical structure governs medical imaging: a pneumonia detector trained across hospitals latches onto the X-ray machine's metadata or image-edge tokens that correlate with the sicker hospital's higher base rate, rather than the lung pathology; it collapses at a new hospital, and the fix is multi-site training that varies the scanner independently of the diagnosis.
Verification¶
Does it exist? Confirmed. This work's DOI resolves to a registered record, which fixes its identity. That is all it fixes.
Does it back the claim? Read against the text for 3 of 4 citations: 1 supported, 2 supported in part. Each verdict is shown under its citation below, with what in the work backs the sentence.
Support is checked per citation rather than per work — the same source can be cited soundly in one article and wrongly in another. Per-citation recording began recently, so a citation with no recorded check is a gap in the record rather than evidence it went unchecked.
See how references were verified.
Registry ID ref:56f02acd9ad6 · see in the full table