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

Area of Science:

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

Background:

• Vision-language models (VLMs) show promise for augmenting radiologic decision support.
• Security vulnerabilities of VLMs, particularly those with optical character recognition (OCR) capabilities, are not well understood in diagnostic workflows.
• Adversarial manipulation of image-embedded text poses a significant risk to diagnostic integrity.

Purpose of the Study:

• To evaluate the security implications of deploying OCR-capable VLMs in radiology.
• To assess the vulnerability of VLMs to adversarial attacks via embedded text.
• To investigate 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 experimental conditions.

Main Results:

• Baseline VLM performance was heterogeneous (median accuracy 0.69).
• Visible and stealth OCR injection significantly degraded performance, causing specificity collapse (FPR 1.00) and reduced accuracy (median 0.43).
• Immune prompting offered only partial mitigation, with residual overcalling and persistent high false-positive rates.

Conclusions:

• Commercial VLMs exhibit a critical failure mode where OCR-readable text can override image evidence.
• Prompt-level defenses are insufficient to protect against these adversarial attacks.
• System-level safeguards, including OCR-aware input handling and human verification, are essential for safe clinical integration of VLMs.