Uncertainty Quantification: A Way Towards Trustworthy Deep Learning in Medical Imaging

In recent years, deep learning has shown impressive performance in medical imaging analysis. However, for a model to be useful in the real world, it needs to be valid and trustworthy. The uncertainty quantification (UQ) methods determine the calibrated level of a model’s confidence in its predictions. Moreover, UQ can demonstrate biases caused by overconfidence or lack of confidence in the model’s predictions. By enabling UQ in medical deep learning models, users can be alerted when a model does not have enough information to make a decision. Consequently, a medical expert could reevaluate the uncertain cases, which would eventually lead to gaining more trust in the model.

Objectives

• Discuss the relationship between output, probability, and uncertainty of predictive models
• Identify the various types of uncertainty that affect model predictions and underline its pivotal role in the outcomes of predictive models
• Define the broad categories of UQ techniques and gain insights into use cases of UQ
• Identify the challenges involved in performing UQ and explore potential solutions

Watch for Free without credit on YouTube!

Additional Resources

• Publication: Quantifying Uncertainty in Deep Learning of Radiologic Images
• Google Colab Notebook: Uncertainty Quantification

CE Credit

SIIM has approved this activity for 1.0 hours of SIIM IIP Credits towards certification and re-certification by the American Board of Imaging Informatics (ABII).