Mitigating Bias in Radiology Machine Learning: 2. Model Development

Area of Science:

• Medical Imaging and Artificial Intelligence
• Radiology AI Development

Background:

• Growing concerns regarding bias and fairness in clinical artificial intelligence (AI) applications.
• Model development is a critical stage in implementing machine learning (ML) tools, susceptible to various biases.

Purpose of the Study:

• To identify and address potential biases in AI model development within radiology.
• To highlight key areas in the model development pipeline where bias can be introduced.

Main Methods:

• Focus on four critical aspects of AI model development: data augmentation, model and loss function selection, optimizer choice, and transfer learning.
• Review of current practices and potential pitfalls in these development stages.

Main Results:

• Bias can be introduced through data augmentation techniques.
• Model architecture, loss functions, optimizers, and transfer learning strategies significantly impact AI model fairness.
• Specific considerations are needed for each stage to prevent bias.

Conclusions:

• Implementing AI in radiology requires careful attention to model development to ensure fairness.
• Adopting appropriate practices during data augmentation, model design, optimization, and transfer learning can mitigate bias.
• Proactive bias mitigation is essential for trustworthy AI in clinical radiology.