Uncertainty Quantification in Medical Image Analysis

An official MICCAI 2023 Tutorial

8th of October 2023 (Sunday AM)

About

This MICCAI 2023 is a half-day tutorial that aims to foster discussions within the scientific community regarding the trustworthiness of deep learning (DL) model predictions by providing valuable insights into uncertainty quantification and model calibration techniques necessary for making reliable decisions. The tutorial combines talks from experts in the field with hands-on sessions.

Participants will gain a comprehensive understanding of uncertainty quantification, covering mathematical foundations and cutting-edge research questions while acquiring practical tools for deriving state-of-the-art uncertainty estimates in medical image analysis.

We believe this tutorial is an essential resource for scientists, clinicians, and industry professionals aiming to enhance the robustness and reliability of their DL models.

Learning objectives

This tutorial will enable participants to gain knowledge about topics ranging from the very fundamentals of uncertainty quantification to cutting-edge recent methods in the field. We expect tutorial attendees to develop:

• A solid understanding of the basic concepts around uncertainty quantification in machine learning and of a taxonomical map of the field.

• The ability to discriminate between different sources of uncertainty.

• Mathematical fundamentals for understanding and implementing several well-established uncertainty quantification techniques.

• Knowledge of the tasks of robustness to and detection of Out-of-Distribution data and Model Calibration.

• Modeling and leveraging multiple annotations for uncertainty quantification.

• Knowledge of quantification beyond classification or segmentation, with basics on conformal predictions.

Program

Part 1: Introduction to Uncertainty, Distribution Shifts and Robustness

by Andrey Malinin

Problem statement: sources of errors and uncertainty, distributional shift
Overview of passive and active solutions
Taxonomy, common confusions, ambiguities and misunderstandings

Part 2: Uncertainty Quantification Techniques and Hands-on

by Vatsal Raina and Nataliia Molchanova

Bayesian perspective on uncertianty quantification (UQ)
UQ methods in medical image analysis
Practical hands-on session
Applications of UQ for real-wold tasks

Part 3: Model Calibration Techniques and Hands-on

by Adrian Galdran and Meritxell Riera i Marín

Calibration: what, when and why?
Visualising and measuring calibration
Improving calibration
Practical hands-on session

Keynote 1: Uncertainty Quantification in Segmentation & Image Reconstruction

Christian Baumgartner

leads the research group for Machine Learning in Medical Image Analysis at the University of Tübingen. He previously worked at ETH Zürich, Imperial College London and King’s College London. His research focuses on developing methodologies to bridge the gap between ML theory and clinical applications. He has a special interest in safety and uncertainty modeling for medical image analysis, and he is an active member of the MICCAI community, having co-founded the UNSURE workshop.

Keynote 2: A Gentle Introduction to Conformal Prediction

Anastasios Angelopoulos

is a machine learning researcher at University of California, Berkeley, working under the supervision of Michael I. Jordan and Jitendra Malik. He works on theoretical aspects of machine learning with applications in vision and healthcare, with a focus on applying modern statistical ideas to increase robustness of black-box models like deep neural networks.