Artificial intelligence and machine learning (AI/ML) technologies are constantly finding new applications across multiple fields. Medicine is no exception, with AI/ML being used for diagnosis, prognosis, risk assessment, and treatment response assessment in a wide variety of diseases. In particular, AI/ML models are finding more and more applications in the analysis of medical images. This includes X-rays, computed tomography, and magnetic resonance imaging. A key requirement for successful implementation of AI/ML models in medical imaging is ensuring proper design, training, and use. In practice, however, it is very difficult to develop an AI/ML model that works well for all members of the population and can be generalized to all situations.
Just like humans, AI/ML models can be biased, which can lead to different treatments for medically similar cases. Despite the factors associated with introducing such biases, it is important to address them and ensure fairness, impartiality, and trust in AI/ML for medical imaging. This requires identifying sources of biases that may exist in medical imaging AI/ML and developing strategies to mitigate them. Failure to do so may result in differing patient benefits and exacerbate inequities in healthcare access.
Professionals from the Medical Imaging and Data Resource Center (MIDRC), including medical physicists, AI/ML researchers, statisticians, physicians, and regulatory agency scientists, as reported in the Journal of Medical Imaging (JMI) This concern has been addressed, being a multi-agency team at home. This comprehensive report covers the five major steps that can occur along the five major steps of medical imaging AI/ML development and implementation, from data collection, data preparation and annotation, model development, model evaluation, and model deployment. It identifies 29 potential sources of potential bias. We have identified biases that can occur in multiple steps. Strategies to reduce bias are discussed and information is also available on his MIDRC website.
One of the main sources of bias is data collection. For example, sourcing images from a single hospital or single type of scanner can introduce bias in data collection. Data collection biases can also arise due to differences in how particular social groups are treated, both during research and across the healthcare system. Additionally, data can become outdated as medical knowledge and practice evolves. This introduces a temporary bias into AI/ML models trained on such data.
Other sources of bias are in data preparation and annotation, which are closely related to data collection. This step allows you to introduce biases based on how your data is labeled before being fed into your AI/ML model for training. Such biases can result from the annotator’s personal biases or oversights related to how the data itself is presented to the user responsible for labeling.
Biases can also arise during model development based on how the AI/ML models themselves are inferred and created. One example is inherited bias, which occurs when the output of a biased AI/ML model is used to train another model. Other examples of biases in model development include biases caused by unequal representation of the target population, or biases resulting from historical circumstances such as social and institutional biases that lead to discriminatory practices.
Model evaluation can also be a potential source of bias. For example, when testing model performance, bias can be introduced by using an already biased dataset for benchmarking or by using an inappropriate statistical model.
Finally, when deploying AI/ML models in real-world settings, biases can also be introduced, primarily from users of the system. For example, bias is introduced when the model is not used for the intended category of image or composition, or when users become overly reliant on automation.
In addition to identifying and thoroughly explaining these potential sources of bias, the team suggests possible ways to mitigate them and best practices for implementing medical imaging AI/ML models. To do. This article therefore provides researchers, clinicians, and the general public with valuable insight into the limitations of AI/ML in medical imaging and a roadmap for remediation in the near future. This may promote fairer and fairer deployment of medical imaging AI/ML models in the future.
Gold Open Access article by K. Drukker et al., “Towards fairness in artificial intelligence for medical image analysis: identifying and mitigating potential biases in the roadmap from data collection to model deployment,” J. Med. image. 10(6), 061104 (2023), Doi 10.1117/1.JMI.10.6.061104.
