09:45 – 10:00 Opening
10:00 – 11:00 Keynote Paul F. Jaeger
Semantic Segmentation and Beyond
11:00 – 11:30 Coffee Break
11:30 – 12:30 Robotic Scene Segmentation
12:30 – 13:30 Surgical Workflow Analysis in the SensorOR
13:30 – 15:00 Lunch
15:00 – 16:30 Gastrointestinal Image ANAlysis (GIANA)
16:30 - 17:00 Coffee Break
17:00 - 18:00 Cataracts
18:00 – 18:15 Closing
Keynote "Semantic Segmentation and Beyond"
Since the publication of the U-Net in 2015, multiple modifications to the architecture have been proposed, suggesting improvements on few or even single data sets. The generalization capabilities of these architectures is questionable, especially given that this year’s Medical Segmentation Decathlon (a challenge which assesses segmentation performance over 10 independent data sets) was so far dominated by the “nnU-Net”, a comparably simple U-Net architecture, that focused on an elaborate training framework instead. Based on the assumption, that semantic segmentation is not the bottleneck of current research in medical image analysis, this talk presents recent work addressing questions beyond segmentation, like how to work with limited training data and how to deal with ambiguous predictions for objects of interest or ambiguous reference annotations.
Paul F. Jaeger is a PhD Candidate at the German Cancer Research Center in Heidelberg, working on detection in medical images. He is co-founder of heidelberg.ai, a community of more than 700 machine learning enthusiasts gathering in monthly events and lectures. He is also part of the so far highest scoring team at this year’s Medical Segmentation Decathlon Challenge.