Pixels and Pitfalls: Building Robust Artificial Intelligence for Medical Imaging

AbstractArtificial intelligence (AI) applications in medical imaging continue to evolve rapidly, with models now capable of interpreting medical images without being trained on explicit labels. This Perspective, based on a conversation with Dr. Pranav Rajpurkar on NEJM AI Grand Rounds,¹ discusses the progression of AI imaging models, starting from early successes in radiology, such as CheXNet, to more sophisticated recent models such as CheXzero. Dr. Rajpurkar emphasizes the importance of understanding the "data generation process," including the artifacts and biases baked into data, which is illustrated by a specific example where an AI model exploited metadata rather than clinically relevant features. Dr. Rajpurkar addresses the urgent need for more open and accessible medical data with his initiative on Medical AI Data for All (MAIDA). We also examine the changing role of clinicians in an AI-augmented health care system, and discuss a collaborative approach where human expertise guides AI development and implementation. Looking ahead, we envision a future where AI systems generate comprehensive medical reports and engage in natural language interactions, while emphasizing the need for ongoing focus on safety, efficacy, and equitable access.