CVPR Daily - Tuesday

Ender Konukoglu , from ETH Zurich , gave an overview of robust and trustworthy AI for medical imaging. He revealed a few key traits of trustworthy AI for medical imaging: 1) robust (adaptable systems robust to changes in input characteristics consistency), 2) interpretable , 3) Self-aware (identifying cases where confident predictions are not possible, 4) Sensitive (useful at realistic operating points that are applicable to clinical use cases). Ben Glocker , from Imperial College London , introduced “safety nets in medical imaging AI” . Safety nets provide a layer of protection against AI failures. In medical imaging AI, we need to make sure that the use of AI is safe and that any predictions made by an algorithm are trustworthy. Especially, Ben discussed various safeguards, including automatic quality control, failure detection, and stress testing, and also delved into robustness and reliability in the context of mismatched data between the method development and clinical deployment. Further, the use of causal reasoning and how it can help identify potential biases concluded his talk. In the first afternoon session, Xiaoxiao Li from the University of British Columbia (UBC ) introduced different advanced federated learning (FL) methods to use more and diverse medical data under privacy regulations, since AI models trained from small data sets are usually not accurate and generalizable. This new training paradigm enables multiple medical institutions to train a model collaboratively without data sharing. In this unique learning regime, Xiaoxiao’s team has investigated novel optimization and learning schemes to tackle data heterogeneity, reduce dependency on data labeling, and adapt FL to different applications. Ismail Ben Ayed from École de Technologie Supérieur (ETS) , introduced how to fully leverage unlabeled data to enhance model generalization in a breadth of real scenarios and applications . Especially, he talked about few- shot learning, unsupervised domain adaptation ad test-time adaptation as a few representative methods. He further introduced a series of latest works that use structure-driven / knowledge-driven / invariance / multi-modal 17 DAILY CVPR Tuesday Medical Computer Vision

RkJQdWJsaXNoZXIy NTc3NzU=