Computer Vision News Computer Vision News 14 physician for correction, ensuring the reliability of the final estimates. Selecting datasets with specific properties conducive to the study’s objectives was a key challenge. Carsten recalls working with a medical dataset focused on lung nodules. “Before we could even start, we looked at different types of uncertainties because you don’t only have one type of uncertainty,” he explains. “You have an uncertainty about the border regions, for example, where everyone agrees it has to be diagnosed, but where they would draw the boundaries differs. Or you have cases distinct from one that the model has already seen, and that’s a different type of uncertainty. How do you create an environment where you can measure both?” Supervisor Paul Jaeger says the question of whether the separation of uncertainty types has any realworld effects has bothered him since he started his PhD in 2016: “I don’t remember being so genuinely curious about the outcome of an analysis before. I am happy the community appreciates our insights and hope they will help to streamline research in the field.” The team used U-Net and HRNet architectures as the computer vision backbone. Within this framework, they define a prediction model, employing techniques such as testtime dropout, incorporating dropout not only during model training but also at test time for uncertainty estimation, and ensembling multiple U-Nets to generate uncertainty estimates over the ensembles of the predictions. Delving into more specific uncertainty measures, the team focused on calculating the uncertainty score at the pixel level. Previous studies typically addressed image classification tasks, where aggregating uncertainty into a single score per image was unnecessary since the analysis didn’t operate at the pixel level. Consequently, these studies lacked pixel-level uncertainty heatmaps and inherently produced uncertainty scores at the image level. “We use predictive entropy, mutual information, and expected entropy between our multiple predictions,” Kim explains. “We have ICLR Oral Presentation
RkJQdWJsaXNoZXIy NTc3NzU=