Bias in artificial intelligence for medical imaging: fundamentals, detection, avoidance, mitigation, challenges, ethics, and prospects

Abstract

Although artificial intelligence (AI) methods hold promise for medical imaging-based prediction tasks, their integration into medical practice may present a double-edged sword due to bias (i.e., systematic errors). AI algorithms have the potential to mitigate cognitive biases in human interpretation, but extensive research has highlighted the tendency of AI systems to internalize biases within their model. This fact, whether intentional or not, may ultimately lead to unintentional consequences in the clinical setting, potentially compromising patient outcomes. This concern is particularly important in medical imaging, where AI has been more progressively and widely embraced than any other medical field. A comprehensive understanding of bias at each stage of the AI pipeline is therefore essential to contribute to developing AI solutions that are not only less biased but also widely applicable. This international collaborative review effort aims to increase awareness within the medical imaging community about the importance of proactively identifying and addressing AI bias to prevent its negative consequences from being realized later. The authors began with the fundamentals of bias by explaining its different definitions and delineating various potential sources. Strategies for detecting and identifying bias were then outlined, followed by a review of techniques for its avoidance and mitigation. Moreover, ethical dimensions, challenges encountered, and prospects were discussed.

Keywords: Artificial intelligence; bias; fairness; machine learning; medical imaging; radiology.

Figure 1

Publication trends about bias in medical imaging artificial intelligence (AI) in comparison with AI in medicine, with different search syntaxes to identify the occurrences of the term “bias” in the title or abstract versus the title alone. Source: PubMed; date of search: May 7, 2024.

Figure 10

Overview of bias avoidance strategies at the data processing phase. Adapted from Gallegos et al. CXR, chest X-ray.

Figure 2

Over-simplified illustration of bias (i.e., systematic error) in contrast to variance, such as random noise.

Figure 3

Main types and sources of bias and related concepts highlighted throughout this review. For other common types and sources of bias, please refer to Table 1.

Figure 4

Over-simplified illustration of optimal and poor representation of subgroups, such as gender in this case, and their effect (*) in subsequent modeling. ROC, receiver operating characteristics.

Figure 5

Over-simplified illustration of covariate shift. Distributional differences between training and test sets lead to poor test performance (i.e., poor generalizability) or significant deviation from the learned function.

Figure 6

Potential and practical bias sources relevant to medical imaging artificial intelligence based on data type (i.e., non-pixel and image data). Radiological images belong to chest computed tomography (upper left panel), chest X-ray (upper right panel), and pituitary magnetic resonance imaging (lower panel).

Figure 7

Human bias in the artificial intelligence life cycle.

Figure 8

Over-simplified illustration of true concept drift while adding new data over time, resulting in changes in the relationship of input features and the target variable and ultimately in model behavior. In medical imaging, this may result from, for instance, a change of reference standard (e.g., new guidelines) in determining the target variable or a difference in the distribution of underlying data. It is also possible that such changes, particularly changes in data distribution, may result in virtual drifts with no obvious difference in model behavior.

Figure 9

Key ethical artificial intelligence principles. WHO, World Health Organization.