Motivation and aim

Many state-of-the-art methods for medical imaging rely on deep learning models that are susceptible to distribution shifts. Several factors cause changes in data acquisition, including ever-evolving scanning technologies and the presence of image artefacts. Likewise, naturally occurring shifts in disease expression and spread can cause the annotated training base to become outdated. As a result, deep learning models deteriorate over time until they are no longer helpful to the clinician.

To maintain the expected performance, models must be updated to incorporate new data patterns while preserving their proficiency in the original evaluation set. We refer to this as lifelong or continual learning. While such a workflow may seem reasonable, actually building, approving and deploying medical lifelong learning solutions faces several challenges.

Our aim with this tutorial is to give participants hands-on insights into how various domain shifts affect the performance of deep learning models in dynamic environments; and help them develop strategies to address these issues and correctly monitor performance. We hereby seek to breach the gap in the MICCAI community between technical research on continual learning and the reality of deploying lifelong learning software in clinics.

Learning objectives

We will take a holistic look at the process of developing and releasing lifelong learning solutions, addressing the following topics:

  • Data drift in medical imaging: Common sources of domain shift and their effect on model performance
  • Dynamic evaluation and monitoring: Metrics for quantifying the performance of lifelong learning models, such as backward and forward transfer, model capacity and computational efficiency; and how to select appropriate test set(s).
  • Continual learning strategies: State-of-the-art methods and how to select the appropriate strategy considering performance, flexibility and resource use.
  • Current regulations for updating models in different global regions.

Schedule

We will explore our learning objectives with expert presentations and in a practical setting. Participants will be divided into small groups. Each group will be assigned a medical imaging task, such as lung nodule detection or cardiac MR segmentation, alongside an existing architecture that achieves good performance in a static setting.

The tutorial will take you through the development of a continual learning product for the given task through three stages:

  1. Identifying relevant domain shifts,
  2. adapting deep learning models for continual learning, and
  3. preparing a product release considering regulations and lifecycle monitoring.