Related Experiment Video
Updated: Jun 13, 2026
09:37
Combining Reflectance Confocal Microscopy with Optical Coherence Tomography for Noninvasive Diagnosis of Skin Cancers via Image Acquisition
Published on: August 18, 2022
AI in Dermato-Oncology: Diagnostic Performance and Prompt-Injection Vulnerability of Vision-Language Models in
Ibrahim Güler1,2, Armin Kraus1, Gerrit Grieb3,4
1Department of Plastic, Aesthetic and Hand Surgery, Otto-von-Guericke University, 39120 Magdeburg, Germany.
Cancers
|June 12, 2026
Summary
Vision-language models (VLMs) show poor accuracy in diagnosing skin cancer from dermoscopic images and are highly vulnerable to adversarial attacks, making them unsuitable for unsupervised clinical use.
Area of Science:
- Dermato-oncology
- Artificial Intelligence in Medicine
- Computer Vision
Background:
- Accurate differentiation of benign melanocytic nevi from invasive melanoma is critical for patient management.
- Vision-language models (VLMs) are being explored for skin cancer assessment, but their reliability needs evaluation.
- Robustness to adversarial manipulation is a key concern for clinical AI deployment.
Purpose of the Study:
- To evaluate the diagnostic performance of three contemporary VLMs on dermoscopic images.
- To assess the susceptibility of these VLMs to single-word adversarial input manipulation.
- To determine the suitability of VLMs for clinical use in dermato-oncology.
Main Methods:
- Fifty-two histopathologically confirmed dermoscopic images (26 benign nevi, 26 melanomas) were analyzed.
- Three VLMs (Claude Opus 4.7, Gemini 3.1 Pro, GPT-5.4) were tested under baseline and adversarial conditions.
- Adversarial conditions involved single-word prompt injection (visual overlay, metadata, or both).
Main Results:
- Baseline diagnostic accuracy ranged from 58.3% to 62.2%, with significant bias in one model missing most melanomas.
- Adversarial conditions drastically reduced accuracy to near-zero levels (0.0-1.9%; p < 10-7).
- Models consistently produced incorrect outputs (98-100%) with unchanged confidence levels.
Conclusions:
- Contemporary VLMs exhibit limited diagnostic accuracy for skin cancer assessment.
- These VLMs are highly vulnerable to minimal adversarial input, leading to near-complete diagnostic failure.
- Current VLMs are not suitable for unsupervised clinical use without safeguards and human oversight.
Related Concept Videos
01:30Skin Cancer
Skin cancer is a type of cancer that occurs when there is an abnormal growth of skin cells, usually triggered by damage to the DNA within the skin cells. It is primarily caused by exposure to ultraviolet (UV) radiation from the sun or artificial sources like tanning beds. Skin cancer is the most common type of cancer worldwide, and its incidence continues to rise.
Basal Cell Carcinoma (BCC): BCC is the most common type of skin cancer, accounting for about 80% of cases. It typically develops in...
Basal Cell Carcinoma (BCC): BCC is the most common type of skin cancer, accounting for about 80% of cases. It typically develops in...
01:16Confocal Fluorescence Microscopy
Confocal microscopy is an advanced microscopic technique. The prime advantage of the confocal microscope over other microscopy techniques is its ability to block the out-of-focus light from the illuminated samples using pinholes. It is widely used with fluorescence optics to obtain high-resolution, sharp contrast images. Unlike optical microscopes, confocal microscopes use a focused beam of light laser to scan the entire sample surface at different z-planes. These microscopes are, therefore,...