TL;DR
While vision-language models (VLMs) are increasingly used to detect synthetic medical images, current benchmarks evaluate them on images alone. This paper reveals that when VLMs are given both an image and its accompanying clinical text, they heavily overweight the text context, causing identical images to receive completely opposite authenticity judgments based solely on metadata changes. To fix this, the authors introduce a paired multimodal benchmark and a training-free inference-time mitigation pipeline called Shortcut-Aware Visual Commitment (SAVC) that successfully neutralizes these textual shortcuts.
What Problem Does This Paper Address?
Generative AI can create highly realistic synthetic medical images, raising serious concerns about insurance fraud and diagnostic deception. To counter this, researchers have turned to Vision-Language Models (VLMs) to detect fakes. However, existing evaluations test these models in a vacuum—using image-only inputs.
In real clinical practice, images never exist in isolation; they are tightly coupled with structured medical records, demographics, and provenance metadata. When VLMs process these joint image-record inputs, they exhibit a critical vulnerability called Text-Induced Decision Shift. If the accompanying text hints that an image might be AI-generated (even via ambiguous phrasing), the VLM often ignores the visual evidence entirely and blindly trusts the text, leading to massive false positives or false negatives.
Methods
The authors tackle this issue with a two-pronged approach:
- A Paired Auditing Benchmark: To isolate the effect of text on authenticity judgments, they hold the medical image completely fixed while systematically swapping out a single metadata field:
Source. They test three conditions: a real-world baseline (Source: Hospital), an ambiguous AI tag (Source: AI-preprocessed), and an explicit AI tag (Source: Edited by Nano Banana). - Shortcut-Aware Visual Commitment (SAVC): A novel, four-stage inference-time mitigation pipeline designed to stop VLMs from taking textual shortcuts without requiring expensive model retraining.
- Stage 0a (Detection): An orchestrator LLM flags metadata fields that could bias the judgment.
- Stage 0b (Counterfactual Verification): The pipeline replaces the flagged field with "Unknown" and checks if the VLM's verdict flips, confirming a causal shortcut.
- Stage 1 (Visual Forensics): The VLM evaluates the image alone to establish an uncontaminated visual baseline and confidence score.
- Stage 2 (Deterministic Integration): If a shortcut is confirmed and visual confidence is high, it relies strictly on the visual verdict. If visual confidence is low/medium, it uses the counterfactual-cleaned text input.
Datasets
The benchmark evaluates robustness across three distinct medical imaging datasets spanning different modalities:
- NIH Chest X-ray14 (NIH-CXR14): Public dataset where synthetic features blend into grayscale anatomy.
- ISIC2019: Public dermoscopy dataset featuring rich textures where edits often leave visible artifacts.
- PediCXR: A private, in-house pediatric chest X-ray dataset used to ensure models weren't simply memorizing public test data.
Synthetic counterparts were generated using an LLM-guided edit-verify-refine loop with Gemini-2.5 models and verified by two in-house clinicians.
Key Insights and Ideas
- Text Overrides Vision: VLMs suffer from "blind faith" in textual metadata. An authentic, perfectly real X-ray can be flagged as a fake simply because its metadata contains an AI-related keyword.
- Naïve Prompting Fails: Simply instructing a VLM to "ignore the Source field" in the prompt (prompt-based suppression) is highly ineffective. The models still get distracted by the text shortcut.
- Scale Doesn't Guarantee Robustness: Larger model capacity does not consistently translate to better multimodal robustness. In some cases, smaller models were more resilient to text-induced shifts than their larger counterparts.
- Visual Reasoning Rewriting: In many cases, changing the text metadata didn't just flip the final answer; it actually corrupted the VLM's internal visual checklist reasoning, causing it to "see" artifacts in a real image that weren't actually there.
Key Results
- On authentic images, appending an explicit AI-origin tag (
Source: AI-edited) caused VLM accuracy to drop by an average of 61.1% across all models and datasets. For instance, on NIH-CXR14, MedGemma-27B's accuracy plummeted from 97.6% to 0.0% just from the metadata change. - Direct prompt-based suppression left models highly vulnerable, with accuracy dropping as low as 1.8% on ISIC2019.
- The proposed SAVC pipeline dramatically improved robustness, recovering up to 99.8% accuracy on MedGemma-27B and 100% on Lingshu-7B within the affected subsets, significantly outperforming image-only baselines and prompt suppression.
Limitations and Future Work
- The study focuses on a single, explicit metadata field (
Source) as a controlled probe. Real-world clinical documentation is far more complex, unstructured, and multi-field. - SAVC is conservative: it neutralizes all confirmed shortcuts even if the metadata happens to be correct, potentially throwing away useful clinical signals.
- The synthetic images were generated via a specific pipeline (Gemini-2.5); future work needs to explore how VLMs handle images from a wider variety of generative architectures (e.g., specialized medical GANs or Diffusion Models).
- The authors suggest exploring algorithmic alignment tuning or attention-based regularization during training to complement their inference-time SAVC solution.
Summary updated