The problem: nodule-risk models are usually built for screening, not the real world
Most deep learning tools for pulmonary nodule malignancy risk are trained and validated on lung cancer screening cohorts, where CT protocols are standardized and the population is a defined age and smoking-history bracket. But the majority of pulmonary nodules radiologists actually report are found incidentally — on a chest CT, or even a CT of the abdomen or spine, ordered for an unrelated complaint. Those patients are more varied in age, scanner type, and disease burden than any screening trial, which is exactly where a model trained on narrower data tends to lose accuracy.
A study from researchers at Radboud University Medical Center and collaborators at Maastricht UMC and University Medical Center Utrecht, published in European Radiology, set out to test that generalization gap directly — across three hospitals instead of one.
What the study tested
The team built a retrospective, multicentre case-control dataset of incidental pulmonary nodules, sampled across three size buckets (5-10 mm, 10-15 mm, 15-30 mm) and aiming for roughly one malignant nodule for every two benign ones per bucket, per centre. The final dataset held 269 nodules from 231 patients, of which 89 were malignant — a deliberately cancer-enriched mix built to stress-test discrimination, not to mimic the low real-world prevalence of malignancy in incidental nodules.
Two deep learning models were compared against the Brock model, the widely used clinical risk-prediction score derived from the Pan-Canadian Early Detection of Lung Cancer Study:
- DL1 — a deep learning model trained on screening data, which had previously shown good discrimination on a single-centre dataset of incidental nodules.
- DL2 — an updated version of the same model, retrained on both screening data and clinical (non-screening) data.
The result: both AI models beat Brock, and stayed consistent doing it
Across the pooled multicentre dataset, DL1 reached an AUC of 0.74 and DL2 reached 0.72, both significantly higher than the Brock model's 0.63 (p < 0.01 for both comparisons, DeLong method). At the sensitivity Brock's standard 10% risk threshold produces (77.5%), both deep learning models held 60% specificity versus 44% for Brock — meaning that at the same catch-rate for cancer, the AI models flagged roughly a third fewer benign nodules as suspicious.
| Centre | DL1 AUC | DL2 AUC | Brock AUC |
|---|---|---|---|
| Centre 1 | 0.76 | 0.75 | 0.74 |
| Centre 2 | 0.72 | 0.71 | 0.55 |
| Centre 3 | 0.73 | 0.71 | 0.59 |
That per-centre breakdown is the real finding. Both deep learning models stayed inside a tight 0.71-0.76 band regardless of which hospital's scans they saw. Brock's accuracy, by contrast, swung from 0.74 at the first centre down to 0.55 and 0.59 at the other two — a collapse the authors tie to differences in CT protocols and patient populations between sites. A risk score can look solid in the hospital where it was built and still lose most of its edge somewhere else; in this dataset, the AI models didn't.
Narrower validation looked stronger — and that's the point
It's worth comparing this to an earlier, single-centre validation of essentially the same screening-trained model, published by an overlapping author group in European Radiology. Tested on 408 incidental nodules from one Dutch academic centre, that model reached an AUC of 0.89 overall (0.78 on a nodule-size-matched subset) versus 0.86 (0.58 matched) for Brock — an even larger apparent gap. The authors of that paper also flagged that the model overestimated malignancy probability and would likely need recalibration before clinical use.
The multicentre study's more modest 0.74/0.72 AUCs aren't a step backward — they're a harder, more honest test. A single centre can flatter a model trained partly on similar data; three centres with different scanners and case mixes are a closer proxy for what happens when a tool ships broadly. That the AI models' advantage over Brock survived that harder test, while Brock's own accuracy did not stay consistent, is the more clinically meaningful result. It's also notable that DL2 — retrained on additional clinical data specifically to improve on incidental cases — performed no better than the original screening-trained DL1, suggesting the generalization gain came from the model architecture and screening-scale training data, not from clinical retraining alone.
Where this fits with structured, whole-volume CT reporting
This study is about a standalone risk score, not a deployed reporting product, and its authors are explicit that prospective validation at real-world (lower) cancer prevalence is still needed before clinical use. But the direction of the finding matters for how incidental nodules get handled day to day: an incidental lung nodule is currently caught, described, and risk-stratified only if the reading radiologist notices it, thinks to apply a risk model, and documents a follow-up recommendation — on a CT ordered for something unrelated, under time pressure. A model that keeps its discrimination steady across scanners and patient populations is a better fit for that reality than one that only works well on the data it was built with. That is the same principle behind structured, whole-volume CT reporting: incidental findings, including pulmonary nodules, are surfaced and risk-stratified on every study by default, with xAID's in-house radiologist reviewing every preliminary report before it goes out ready-to-sign.
Frequently asked questions
What did the new study find about AI and incidental lung nodules?
In a multicentre European Radiology study, two deep learning models scored incidental pulmonary nodules with AUCs of 0.74 and 0.72, both significantly better than the 0.63 AUC of the established Brock risk model. At a matched sensitivity of 77.5%, the deep learning models reached 60% specificity versus 44% for Brock.
Did the AI models perform consistently across different hospitals?
Yes. Across the three participating centres, the deep learning models' AUCs stayed in a narrow 0.71-0.76 band. The Brock model's accuracy varied far more by site, ranging from 0.74 at one centre down to 0.55 and 0.59 at the other two, where CT protocols and patient populations differed.
Is an incidental lung nodule the same as one found on lung cancer screening?
No. Screening-detected nodules come from patients already enrolled in a lung cancer screening program using standardized low-dose CT. Incidental nodules turn up unexpectedly on CT scans ordered for an unrelated reason, in patients with far more varied scanners, protocols, ages, and comorbidities — which is why models validated only on screening data don't automatically generalize to incidental findings.
What does this mean for how incidental lung nodules get reported?
It reinforces that malignancy-risk scoring for incidental nodules works best when it happens automatically and consistently, on every chest CT, rather than depending on which radiologist notices a nodule and remembers to apply a risk calculator. Structured, whole-volume CT reporting is built around flagging and risk-stratifying incidental findings like this by default.
Source: Dinnessen R, et al. "Multicentre performance and consistency of two deep learning models for malignancy probability estimation of incidental pulmonary nodules." European Radiology (2026), doi.org/10.1007/s00330-026-12787-y. Prior single-centre validation: doi.org/10.1007/s00330-025-11829-1. Figures are rounded as reported.