Teaching medical AI to stop inventing findings can also expose what it misses

Ask an image-reading AI whether a scan contains a particular finding, and the question itself can become a misleading clue. A model may report what the wording suggests rather than what the image supports. But reducing those unsupported reports is not the same as making the model a better medical reader.

That distinction drives a preprint by Bhaskar Gurram. A small medical vision-language model learned to report far fewer digitally erased lesions than frontier models. Yet on a fresh external test, it missed all 48 images showing pneumothorax, a critical chest finding. The strongest result and the most serious failure belong together.

Change the image, keep the question

The study used public medical datasets containing expert-drawn masks: outlines identifying the pixels belonging to a lesion. Software generated several views of each case. One retained the original image. Another removed the masked lesion using inpainting, which fills the selected region with replacement image content.

A control view applied the same editing tool to a region of the same shape elsewhere, leaving the lesion intact. This matters because a model might otherwise learn that any edited-looking image deserves a negative answer. Unedited lesion-free images and deliberately degraded images provided additional checks.

The main measure, the evidence sensitivity gap, subtracts the rate of absent-or-refer responses on control-edited images from that rate on lesion-erased images. A model that always says no—or reacts identically to both edits—scores zero. A higher score indicates that its answers distinguish the two conditions, although anatomy-dependent editing traces could also contribute.

A large internal improvement

MedScope-4B adapts the four-billion-parameter Qwen3.5-4B model while keeping its image encoder frozen. Training used examples derived from dataset labels, followed by selection of successful model-generated responses and replay of ordinary medical questions. The agent could zoom or adjust image contrast.

On the main internal test of 507 cases across five imaging types, MedScope-4B achieved an evidence sensitivity gap of 0.758. It claimed a lesion remained in 9.6% of lesion-erased views, compared with roughly 41–62% for nine frontier models evaluated through the same agent protocol.

Its covered-and-correct rate on original and control-edited lesion views was 84.9%, versus 82.9–88.8% for those comparators. None of those differences was statistically significant, but that does not establish equivalent accuracy: the uncertainty intervals still allowed meaningful disadvantages.

Training comparisons support a role for the counterfactual images. Across three training seeds, removing them lowered the gap to an average of 0.279 and raised erased-lesion claims to about 68%. Simply matching the proportion of negative training answers did not recover the benefit. The effect also appeared with a second base model and transferred to two editing tools absent from training.

What the experiments did not establish

Control-edit examples helped reduce the shortcut of denying a lesion whenever an image looked edited. However, a further experiment did not establish that precisely matching the control edit’s shape and tool was necessary.

Nor did the recipe establish superior evidence sensitivity on fresh external data over an adapted removal-based preference-training method. That alternative teaches a model which of two answers to prefer. The methods produced different balances of missed lesions and unsupported findings, rather than a universal winner.

Additional paired-view consistency training also failed its pre-specified primary test against a matched-compute control. Longer training alone reduced erased-lesion claims on unseen editing tools at every seed, but reduced real-lesion sensitivity at one.

The failure hidden by aggregate scores

At external centres, frontier models detected more real lesions. On the fresh external test, MedScope-4B’s overall utility under the study’s assumed error costs was below gemini-4-argon’s.

Most strikingly, MedScope-4B confidently answered no on all 48 unedited pneumothorax images. Recovering higher-resolution originals did not fix the final model’s failure.

A post-hoc analysis identified sparse positive training coverage as a plausible explanation, not a demonstrated cause. Stage one contained only four pneumothorax lesion images, while validation contained no positive examples asked about that finding. Later training shifted chest-image answers towards no. The selection procedure therefore could not detect the disappearance of this capability.

Evidence sensitivity is not clinical readiness

The lesion erasures passed automated checks, not completed physician validation; the reader study was still pending. Residual lesions, secondary signs and editing artefacts could affect the expected answers. Some public labels were disputed, and cross-dataset same-patient overlap could not be fully excluded.

The pre-registration also needs context: the main comparator leaderboard was already known when the revised training plan was written, although fresh confirmatory sets were untouched until model selection.

This account uses the supplied paper-text extraction, which omits figures and may damage tables or equations. The work is a retrospective preprint, not prospective clinical validation. Its practical lesson is narrower than a safer medical AI: test whether answers follow visual evidence, but also require validation positives and sensitivity safeguards for each critical finding. Fewer invented lesions cannot compensate for a finding the model never detects.