OCR-mediated modality dominance in vision-language models: implications for radiology AI trustworthiness

Area of Science:

• Artificial Intelligence in Medical Imaging
• Computer Vision Security
• Radiology Decision Support Systems

Background:

• Vision-language models (VLMs) show promise for augmenting radiologic decision-making.
• Security vulnerabilities of VLMs, particularly those with optical character recognition (OCR) capabilities, are not well understood in clinical contexts.
• Image-embedded text can be a vector for adversarial attacks, impacting diagnostic integrity.

Purpose of the Study:

• To evaluate the security implications of deploying OCR-capable VLMs in radiology workflows.
• To assess the susceptibility of VLMs to adversarial text injections, both visible and imperceptible.
• To determine the effectiveness of prompt-level defenses against such attacks.

Main Methods:

• Ten VLMs were tested on 600 brain MRI studies for tumor detection.
• Evaluations included clean input, visible report injection, stealth OCR injection, and immune-prompt defense.
• A total of 27,000 inference calls were analyzed across five conditions.

Main Results:

• Baseline VLM performance was heterogeneous (median accuracy 0.69).
• Visible and stealth OCR injections significantly degraded performance, causing specificity collapse (FPR 1.00) and reduced accuracy (median 0.43).
• Immune prompting offered only partial mitigation, with persistent false positives.

Conclusions:

• OCR-capable VLMs exhibit a critical failure mode where embedded text can override image evidence, even under stealth conditions.
• Prompt-level defenses are insufficient to protect against these vulnerabilities.
• Clinical deployment requires system-level safeguards like OCR-aware input handling and human verification.