47 In the Picture: Medical Imaging Datasets Computer Vision News Computer Vision News The workshop kicked off with a dedicated session, chaired by Amelia, where early career researchers, including PhD students, postdocs, and research assistants, had the opportunity to present their recent works on quality annotations, dataset tracking, citations, and recommendations for data governance in healthcare. Next, Judy tackled the issue of shortcut learning in medical imaging, where AI models often rely on irrelevant features for predictions instead of genuine pathology. This reliance can lead to biased outcomes for certain subgroups, particularly concerning age, gender, and race. Despite their potential, these biases hinder AI’s effectiveness in improving patient care, especially for historically underserved groups, highlighting the need for more robust approaches. Leo and Enzo’s sessions focused around dataset annotations. Leo explored the current state of the field, the development process for dataset creation, shared lessons from common mistakes, and engaged participants in dynamic discussions about critical topics. "Medical imaging data collection, curation, and annotation for the development of a Radiology App is a labor-intensive and laborious process. It is a lengthy process that requires interdisciplinary expertise from junior and senior actors," said Leo. He added: "An iterative bootstrapping approach aims to optimize the expertise of various
RkJQdWJsaXNoZXIy NTc3NzU=