Enhancing Dataset Quality for AI in Radiology: Challenges and Solutions

The quality of datasets used in artificial intelligence (AI) applications in radiology is pivotal for the development of robust and accurate AI models. Various issues impacting dataset quality have been identified, affecting the performance and generalizability of AI tools in clinical settings. First, many datasets lack diversity in demographic, geographic, genetic, and disease representation. This inclusivity deficit can lead to AI models that perform poorly across different populations, potentially exacerbating health disparities.[1] Second, label accuracy in datasets is a significant concern. Studies have shown that labels in some large public datasets do not accurately reflect the visual content of images, leading to AI models trained on incorrect data and thus compromising their reliability.[2]