When the Report Is Present, Is the Medical VLM Still Looking at the Image?

Introduction

Medical vision-language models can process a radiograph together with a clinical question and a radiology report. That combination creates a basic evaluation problem: if the report already states the relevant finding, a model may produce the expected answer without making substantial use of the image. A correct output, in other words, does not necessarily prove that the model inspected the scan.

A study presented through Hugging Face Daily Papers proposes ModaLens, a paired image-swap audit designed to measure this behavior. Its central question is simple: if the question and report stay exactly the same but the image is replaced, does the model’s answer change?

How the audit works

The main experiment evaluates MedGemma-27B on 3,199 paired MIMIC-CXR cases from 293 patients. Each case is tested with 14 questions: 13 finding-specific questions and one composite question. Researchers keep the question and report fixed, then replace the image with one from another study. In most substitutions, the replacement comes from the same patient, which helps reduce confounding from identity and broad clinical history.

The model is evaluated in two settings: with the report available and without it. The researchers measure whether the generated answer changes after the swap, while also examining changes in continuous answer scores that may not flip a binary prediction.

Main findings

• Under an explicit answer instruction, answers changed in 4.26% of image-swap trials when the report was present.
• Without the report, the rate rose to 20.94%, a paired increase of 16.7 percentage points. The patient-clustered 95% confidence interval was 15.6 to 17.7 points.
• A version using the original prompt and a lowercase first-token readout showed the same direction: 4.70% with the report versus 17.07% without it.
• Image substitutions also moved continuous answer scores in cases where the binary answer remained unchanged, suggesting that a simple label-flip metric can miss part of the image’s influence.
• The direction was replicated in two additional model lineages, although the supplied abstract does not report their exact figures.

What the result means

Under this protocol, the presence of a report made the model less sensitive to image substitutions. One interpretation is that the report supplies a strong textual shortcut: once the relevant finding is stated, the model can answer from the text and has less incentive to use the current image. This should not be overstated as proof that the model ignores images altogether, nor as a direct measure of clinical unreliability.

The study’s most important caveat is that its labels were derived from radiology reports. ModaLens therefore measures whether changing the image changes the output; it does not by itself determine whether the model’s visual interpretation is medically correct. Prompt wording, output parsing, model family, and the choice of image pair can also affect the measured sensitivity.

The broader contribution is methodological. Instead of relying only on accuracy, researchers can test whether a multimodal system responds to the evidence source that it is supposed to use. For medical AI, stronger audits will likely need controlled image substitutions, fixed textual context, continuous-output analysis, and independent visual ground truth. These steps can help separate genuine image-based reasoning from answers that are merely consistent with a report.

Source: Hugging Face Daily Papers