Computer Vision News - July 2022

BEST OF CVPR 39 Medical Computer Vision Jie Ying Wu from Vanderbilt University presented “Modeling robotic surgery” in person. With the increased adoption of surgical robots and their ability to record movements and events, Jie Ying introduced how to reconstruct entire surgeries. This, combined with recent advances in machine learning offers an unprecedented opportunity to improve surgical outcomes by modeling the interactions between each part. Finally, Daniel Rückert from TU Munich gave a wonderful talk on “Learning clinically useful information from medical images” . In this talk, Daniel emphasized that current AI solutions are often brittle due to inadequate training data, and there are well-justified privacy concerns. Further, he elaborated on his latest works of building privacy-preserving AI by using federated learning to reduce the need of sharing data, and by providing end-to-end privacy guarantees using cryptography and differential privacy methods . latest work on individualizing encoding models which suggest that neural encoding models need not be population-based, but can also be subject-specific. This finding can bequiteuseful formanyclinical applications. In the final session, Oliver Taubmann from Siemens Healthineers delivered the first talk on “Vision in the CT Workflow” from the industrial perspective, where he emphasized the importance of deploying explainable models such as counterfactual- based decision models for increasing the meaningfulness, intuitiveness, and enhancing imaging quality, which is crucial for real-world CT workflow. Ulas Bagci from the Northwestern University , introduced how to build trustworthy AI for Imaging-based Diagnoses from the following three perspectives: 1) algorithmic robustness , 2) interpretable ML methods and 3) expert-in-the-loop systems . Especially, the combination of different aspects could offer the opportunity to build the most trustworthy medical AI system.