Summary

Have you ever worked with data from five, or even ten years ago? You probably noticed how different it is from more recent cases. If you did not, your model definitely did.

In medical imaging, there are many factors that cause the data distribution to change over time, which is reflected in abrupt performance drops in deep learning models. We may reduce this effect through domain adaptation and improved generalization, but a point will inevitably come where we need to update our model with new data. And when that happens, we may no longer have access to all training data used in the past.

Continual learning addresses precisely this question by developing strategies that acquire new information without losing previous knowledge. This opens up attractive possibilities, such as extending the lifespan of medical software solutions and leveraging large amounts of multi-institutional data without the need for federated protocols. Yet deploying continual learning solutions comes with a number of technical and legal challenges and requires additional quality monitoring.

This tutorial will cover potential sources of domain shift, continual learning metrics and strategies, and regulatory guidelines. You will be assigned a group working on a particular medical imaging task, and each group will take the role of a company looking to release an AI solution that learns over time. By combining expert presentations on key topics and practical assignments where participants can engage with the subject hands-on, we hope to convey the core opportunities and challenges of releasing lifelong learning solutions.


Speakers

Jayashree Kalpathy-Cramer

University of Colorado and Harvard Medical School

Jayashree Kalpathy-Cramer leads the Artificial Medical Intelligence Division at the University of Colorado ophthalmology department. She has previously co-directed the Quantitative Translational Imaging in Medicine lab and the Center for Machine Learning at the Athinoula A. Martinos Center at Harvard Medical School. In addition to developing novel machine learning algorithms, she is actively engaged in applying these to clinical problems in radiology, oncology and ophthalmology. She was funded through NIH to develop quantitative imaging methods in cancer and is the PI of an NSF-funded project to develop and apply algorithms to build diagnostic tools in ophthalmology. Research from this work has resulted in a deep-learning based algorithm for disease diagnosis and response assessment that is currently being evaluated at several clinics and screening trials in the US and India.

Website

Praveer Singh

University of Colorado

Dr. Praveer Singh is an Assistant Professor in the Division of Artificial Medical Intelligence within the Department of Ophthalmology at the University of Colorado School of Medicine. He received a Bachelor's degree in electrical engineering from NIT Kurukshetra in 2012. Subsequently, he pursued a Master's in Telecommunication Engineering at Telecom ParisTech in 2014, and a Ph.D. at the Imagine Lab of Ecole des Ponts ParisTech in 2018. Prior to joining the University of Colorado, Dr. Singh held joint postdoctoral research fellow positions at the Athinoula A. Martinos Center for Biomedical Imaging at Massachusetts General Hospital and Harvard Medical School. He has received notable accolades, including the DAAD-Wise Scholarship for pursuing an internship in Germany, the French Government's Charpak Scholarship, and the Telecom Foundation Scholarship during his Master's studies and a University of Paris-Est Fellowship for pursuing his Ph.D. His pioneering work on unsupervised representation learning by predicting image rotations, popularly known as RotNet, has been highly and widely recognized in the computer vision and machine learning research community. Dr. Singh has authored more than 30 papers in prestigious peer-reviewed journals and conference proceedings and is the inventor of two US patents. His research endeavors revolve around the overarching goal of developing unbiased, robust, and generalizable AI tools and algorithms with a special focus on translation these advanced technologies into clinical applications, notably in ophthalmology, radiology and neonatology.

Website

Martin Mundt

Technical University of Darmstadt

Martin Mundt is a hessian.AI junior research group leader of the Open World Lifelong Learning lab, where the focus is to create robust systems that learn continually in an open-ended world. He is also a board member of directors at the non-profit ContinualAI organization for the 2022-2024 election term. He obtained his PhD in continual deep learning at the Goethe University Frankfurt, for which he received the best thesis in natural sciences award.

Website

Organization

Camila Gonzalez

Stanford University

Website

Anirban Mukhopadhyay

Technical University of Darmstadt

Website

Maria Zuluaga

EURECOM, 3IA Institute Côte d'Azur and King's College London

Website

Ilkay Öksüz

Istanbul Technical University and King's College London

Website

Niklas Babendererde

Technical University of Darmstadt

Website

Khrystyna Faryna

Radboud UMC

Website

Magda Paschali

Stanford University

Website

John Kalkhof

Technical University of Darmstadt

Website

Moritz Fuchs

Technical University of Darmstadt

Website