Abstract
The rapid advancement of foundation models in medical imaging represents a significant leap toward enhancing diagnostic accuracy and personalized treatment. However, the deployment of foundation models in healthcare necessitates a rigorous examination of their trustworthiness, encompassing privacy, robustness, reliability, explainability, and fairness. The current body of survey literature on foundation models in medical imaging reveals considerable gaps, particularly in the area of trustworthiness. Additionally, extant surveys on the trustworthiness of foundation models fail to address their specific variations and applications within the medical imaging domain. This survey paper reviews the current research on foundation models in the major medical imaging applications, with a focus on segmentation, medical report generation, medical question and answering (Q&A), and disease diagnosis, which includes trustworthiness discussion in their manuscripts. We explore the complex challenges of making foundation models for medical image analysis trustworthy, associated with each application, and summarize the current concerns and strategies to enhance trustworthiness. Furthermore, we explore the future promises of these models in revolutionizing patient care. Our analysis underscores the imperative for advancing towards trustworthy AI in medical image analysis, advocating for a balanced approach that fosters innovation while ensuring ethical and equitable healthcare delivery.
Introduction
Foundation models (FMs) such as the Segment Anything Model (SAM) and GPT-4 are revolutionizing medical artificial intelligence. However, their "black-box" nature poses significant risks in high-stakes clinical environments. While general AI surveys exist, they often overlook the unique constraints of medical imaging, such as data scarcity, strict privacy laws, and the requirement for extreme precision. This paper provides a comprehensive survey of FMs in medical imaging through the lens of trustworthiness, establishing a novel taxonomy across five critical dimensions: privacy, robustness, reliability, explainability, and fairness.
Methodology and Taxonomy
This survey systematically categorizes 31 recent foundation models and 76 research papers. The landscape is organized into two primary axes: adaptation strategies (how models are used, e.g., fine-tuning vs. off-the-shelf) and trustworthiness pillars.
The core of the survey is built upon five Trustworthiness Pillars:
- Privacy (T1): Implementation of techniques like Federated Learning and Differential Privacy to protect sensitive patient data.
- Robustness (T2): Resilience against noisy inputs, distribution shifts, and adversarial attacks.
- Reliability (T3): Addressing "hallucinations" and miscalibrated confidence scores.
- Explainability (T4): Making model decisions interpretable through heatmaps or concept-bottlenecks.
- Fairness (T5): Ensuring equitable performance across demographics such as race, gender, and BMI.
Clinical Insights and Challenges
Benchmarking Trust
The paper reviews a vast array of datasets, highlighting specific "trust-focused" benchmarks. These include Harvard-FairVLMed for evaluating fairness in vision-language models and Med-HallMark, a benchmark specifically designed to detect hallucinations in medical FMs.
The "SAM" Reality Check
While SAM is a powerful tool for general segmentation, it is not a universal solution for medical imaging. Research indicates significant performance drops on specific medical modalities. Furthermore, SAM is highly sensitive to "prompt jitter," where small variations in box or point prompts lead to inconsistent segmentation results.
Hallucinations and Fairness
A critical insight in medical report generation is the tendency of models to hallucinate references to prior reports. This occurs because models are trained on real clinical text containing such references, yet they lack the historical context of the specific patient.
Fairness also remains a concern; for instance, the MedSAM model shows a statistically significant negative correlation between patient BMI and segmentation performance. Additionally, performance disparities have been noted between male and female patients in certain diagnostic contexts.
Key Quantitative Findings
- Federated Learning: One study showed that parameter-efficient federated learning on SAM resulted in a 6% higher Dice score for kidney CT segmentation compared to standard approaches.
- Vulnerabilities: CLIP-based medical models are susceptible to "backdoor attacks," where poisoned data causes predictable failures on specific triggers.
- Privacy Trade-offs: While Differential Privacy (DP) is essential for data protection, it often leads to performance degradation, presenting an ongoing challenge for researchers.
Future Directions
The authors identify several frontiers for future research:
- Annotation Bottleneck: Scaling techniques like Reinforcement Learning from Human Feedback (RLHF) is difficult due to the high cost of expert radiologist time.
- Interdisciplinary Alignment: There is a need to bridge the gap between technical definitions of robustness and clinical definitions of safety.
- Regulatory Frameworks: Ethical and regulatory guidelines must evolve to keep pace with multi-modal foundation models.
Ultimately, the path toward clinically deployable foundation models requires a synthesis of computer science, ethics, and clinical medicine to ensure that innovation is balanced with ethical and equitable healthcare delivery.
Summary updated