Keynote 1

Title: nnU-Net - the baseline, the tool, the framework. Insights, analysis and future directions

Name: Fabian Isensee

Affiliation: German Cancer Research Center

Abstract: Semantic segmentation is a ubiquitous problem in biomedical image analysis. One of the most challenging problems in this domain are the diverse properties that our datasets manifest: size, spacing, anisotropy, modality, ... . As a consequence, methods poorly transfer between datasets and often require painful manual adaptation. In this context, nnU-Net has stood out as an out-of-the box segmentation tool not only by adapting to each individual dataset fully automatically but also by continuously challenging segmentation quality of the most sophisticated competing methods. Since its conceptualization in 2018, nnU-Net has withstood the test of time, still delivering state-of-the-art performance today. In this presentation we will revisit nnU-Net's core idea and implementation as well as shed light on its conceptualization, derive potential explanations for its performance and formulate possible directions for improving it looking forward.

Keynote 2

Title: Data, Assemble: Towards Efficient Medical Image Analysis

Name: Zongwei Zhou

Affiliation: Johns Hopkins University

Abstract: The success of deep learning relies heavily on large labeled datasets, but we often only have access to small datasets with partial labels. For the development and evaluation of deep learning algorithms, there is a critical need to build large-scale, multi-center, fully-labeled datasets in an annotation-efficient manner. In this talk, I will introduce a new initiative—"data, assemble"—that assembles large-scale datasets from available data resources, instead of piling up data and labels from scratch for every task, overcoming the deficiency of the current annotation paradigms. In particular, I will dive into our recent discovery: Learning from the classes in "negative examples" can better delimit the decision boundary of the class of interest. This discovery is the foundation of the "data, assemble" initiative, underlining the necessity of assembling multiple datasets with diverse (yet partial) labels. It also sheds new light on the computer-aided diagnosis of rare diseases and emerging pandemics, wherein "positive examples" are hard to collect, yet "negative examples" are relatively easier to assemble.