Ethical considerations, inclusion, biases and data protection for augmented intelligence in Medical Imaging in the era of Generative AI

In the era of generative artificial intelligence (GAI), ethical considerations, inclusion, biases, and data protection have become paramount, especially in the realm of medical imaging. While GAI's deep learning models have significantly advanced diagnostic accuracy, their inherent 'black box' nature presents a transparency challenge. As these models are complex and often uninterpretable, performance metrics such as Receiver Operating Characteristic (ROC) are vital in assessing system performance and building clinicians' trust. We also bring attention to the issue of automation bias, where an over-reliance on AI systems without critical evaluation can lead to misdiagnosis. A proposed solution is the integration of a human-in-the-loop approach, especially with practitioners with less experience.

The implementation of GAI also necessitates a robust framework that respects ethical principles such as beneficence, justice, and respect for persons, to ensure data protection and ethical use of patient data. A significant concern is the lack of diversity in training models, which can exacerbate health disparities. For instance, a deep learning model used in cardiac imaging was found to be less accurate in diagnosing minority ethnic groups due to a lack of representation in the training dataset. We underscore the need for diverse and inclusive training datasets in the GAI era, to ensure equitable healthcare outcomes.