Advances in Uncertainty Quantification for Deep Learning-based Medical Image Analysis

Join us on Feb 13 at 4:15 PM in Stevenson 5326 for “Advances in Uncertainty Quantification for Deep Learning-based Medical Image Analysis,” presented by Brayden Schott, a graduate student at the University of Wisconsin School of Medicine and Public Health!

Date: Feb 13

Time: 4:15pm

Location: Stevenson 5326

Speaker: Brayden Schott

Bio:
Brayden Schott is a graduate student at the University of Wisconsin School of Medicine and Public Heath, in the Department of Medical Physics. He currently studies deep learning and uncertainty quantification, as applied to oncology and medical imaging.

Title:
Advances in Uncertainty Quantification for Deep Learning-based Medical Image Analysis

Abstract:
The predictive outputs of deep learning models inherently contain uncertainty, which is often imperceptible to the user. Without uncertainty information, the reliability of model predictions cannot be ensured, and model failures may not be appropriately accounted for. This poses a threat within clinical settings, where deep learning models are increasingly being studied with the intent to inform patient care. Thus, uncertainty quantification methods for clinical models are essential yet remain under-explored.

In this seminar, we will examine our latest contributions to uncertainty quantification for medical image analysis tasks. We will first discuss our approach to an out-of-distribution detection algorithm which incorporates information bottleneck optimization to more consistently flag images that differ from those present in a model’s training set. We will then discuss our exploration of leveraging a trained model’s gradient space to capture predictive uncertainty. Along the way, we will demonstrate the utility of these approaches, address their challenges, and discuss their potential to foster the safe and responsible deployment of clinical deep learning models.

Explore Story Topics