CARES: A Comprehensive Benchmark of Trustworthiness in Medical Vision Language Models

Published: arXiv, 2024
Link: arXiv:2406.06007v1
Summary: CARES evaluates the trustworthiness of medical vision language models (Med-LVLMs) across five dimensions: trustfulness, fairness, safety, privacy, and robustness.

Problem

Unverified trustworthiness of Med-LVLMs poses risks in medical applications.
- Factual inaccuracies in medical diagnoses.
- Overconfidence in generated diagnoses.
- Privacy breaches.
- Health disparities across demographic groups.

Contributions

Introduction of CARES benchmark for evaluating Med-LVLMs' trustworthiness.
Assessment across five critical dimensions: trustfulness, fairness, safety, privacy, and robustness.
Public release of benchmark and code.

Method

Dataset from seven medical multimodal and image classification datasets.
18K images and 41K question-answer pairs in various formats.
Evaluation based on trustfulness, fairness, safety, privacy, and robustness.

Result

Models show factual inaccuracies.
Poor uncertainty estimation.
Models fail to maintain fairness across different demographic groups.
Vulnerability to attacks.
Privacy leaks.
Inadequate handling of OOD samples.

Conclusion

Existing Med-LVLMs are unreliable and pose significant trustworthiness issues.
CARES aims to drive standardization and development of more reliable Med-LVLMs.

Reference

Peng Xia, Ze Chen, Juanxi Tian, Yangrui Gong, et al. "CARES: A Comprehensive Benchmark of Trustworthiness in Medical Vision Language Models." arXiv preprint arXiv:2406.06007v1, 2024.