Two medical AI models missed every collapsed lung on chest X-rays

An audit of five open medical image-reading models finds that a single accuracy score can hide total failure on one diagnosis. Two models never flagged a pneumothorax, yet explained their wrong answers fluently.

5 min readIIBenchmarkedPreprint, not peer reviewed
Save
A photograph of an old wooden sorting tray on a warm off-white paper surface, with its compartments neatly filled with small seeds, pebbles, and buttons, except for one empty compartment, and a small deflated deep blue balloon sitting neglected on the paper beside it.AI illustration
Illustration: AI generated

What the study shows

The authors tested five open medical vision-language models. These are AI systems that look at an image and answer in text. Each model answered 6,048 yes/no questions about chest X-rays from a public dataset called ChestX-ray14. Each of seven findings had 432 X-rays where it was present and 432 where it was absent. The paper is a preprint accepted at a NeurIPS 2026 workshop and has not been peer reviewed.

The headline result concerns pneumothorax, a collapsed lung that can kill within hours if untreated. Two models, HuatuoGPT-Vision-7B and Qwen2.5-VL-7B, answered "no" to every pneumothorax question. On the same images, Lingshu-7B found about two thirds of cases and MedGemma-4B found 38%. The failing models still wrote specific explanations, describing the lungs as fully inflated, with no hedging.

Overall accuracy did not show this. MedGemma scored highest overall, about 70% against a coin-flip level of 50%. Yet it had its own problem: it caught 95% of atelectasis, a partly collapsed lung, but wrongly flagged about two thirds of healthy cases.

The authors report the pneumothorax failure held across four ways of asking the question. It held at full image resolution and on X-rays showing only pneumothorax. It also repeated on a second dataset, SIIM-ACR, where radiologists had outlined each collapsed lung by hand. There, the failing models found none of 432 cases, including none of the 144 largest.

A tenfold larger Qwen model, at 72 billion parameters, caught only about 4% of pneumothorax cases. MedGemma did improve at its larger size, so scale helped one model family but not the other.

The authors also looked inside the models. The failing models were highly confident in their wrong "no" answers. Yet a simple classifier trained on their internal signals, called a linear probe, could partly tell pneumothorax cases apart. The authors read this as the finding being registered inside the model but lost before the answer.

Our reading

In our reading, the strongest part of this paper is the evaluation design, not the mechanism. Balancing every finding, age band and sex means a model cannot look good by leaning on a common diagnosis. That design made the collapse visible. The second dataset with radiologist outlines answers the obvious worry that the original labels were wrong.

The practical lesson is simple. A model can score respectably on average and never make one diagnosis at all. Worse, its explanation reads like a careful negative. Imagine a triage tool built on Qwen2.5-VL-7B sorting overnight X-rays. A collapsed lung would be marked "no" with a tidy rationale, and nothing in the output would prompt a second look.

We rate this work at level II, Benchmarked. It reports results on public datasets for named open models. It does not reach level III, because the authors have not released their code or their list of test items.

Free membership

Read the rest of this story

Membership is free. You get every article in full and the Issue every Sunday.

We send a sign-in link, no password needed. Already a member? Sign in

Readiness assessment

II
Benchmarked
LevelStatusEvidence
I  Lab onlyMetThe work is an evaluation study on research datasets with no deployed system.
II  BenchmarkedMetResults are reported on public datasets (ChestX-ray14 and SIIM-ACR) with defined metrics for five named open models, though the evaluation set is the authors' own construction and not a standard benchmark.
III  Open sourceNot yetThe paper states no code is released and the item manifest and scripts are not released. The datasets and models are public, so the set can be rebuilt from the description. No link is given.
IV  In productionNot yetThere is no documented use by a named organisation. The paper is itself an audit that argues these models should not be trusted for per-finding reading without checks.

What would move it up
Releasing the evaluation code and the item list would reach level 3 as a reproducible audit. Independent replication by another group, on other findings and other models, and a test of whether an output-level fix recovers detection, would strengthen it. Peer review is also still needed. Level 4 would need a named organisation using the audit in a real model-selection or safety process.

Checked automatically every week, last on 16 November.

The study

Shreya Ajay Kale, Angad Ripudaman Singh Bajwa. Aggregate Accuracy Hides Diagnosis-Specific Failure in Medical Vision-Language Models. arXiv, 11 November 2026. Read the paper

Preprint, not peer reviewed · Licence arXiv (no reuse)

Notes

  1. Kale and Bajwa, Aggregate Accuracy Hides Diagnosis-Specific Failure in Medical Vision-Language Models. arXiv preprint, November 2026. NeurIPS 2026 Med-Reasoner workshop. Not peer reviewed. Source
  2. No code released. The paper's checklist states the item manifest and generation scripts are not released.
  3. Pneumothorax recall on ChestX-ray14 (432 positives): MedGemma-4B 0.377, Lingshu-7B 0.671, HuatuoGPT-Vision-7B 0.000, Qwen2.5-VL-7B 0.000 (Table 2).
  4. Overall accuracy on 6,048 items: MedGemma 0.698, Lingshu 0.620, HuatuoGPT 0.534, Qwen 0.518, MedVLM-R1 0.502 (Table 1).
  5. SIIM-ACR pneumothorax recall: Lingshu 0.861, MedGemma 0.470, HuatuoGPT, Qwen and MedVLM-R1 0.000 (Table 3). Source
  6. Qwen2.5-VL-72B pneumothorax recall 0.044 (19 of 432). MedGemma rose from 0.377 at 4B to 0.565 at 27B.
  7. Qwen pneumothorax confidence 97.6% at 50.0% accuracy. Linear-probe AUC on SIIM-ACR: Qwen 0.716, HuatuoGPT 0.746, Lingshu 0.851, MedGemma 0.891 (Tables 4 and 5).

Edited and signed by

Stefan PerdalEditor and publisher86 stories edited

Stefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.

How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

More on Safety & evaluation