Computer Vision News Computer Vision News 48 Workshop actors so as to minimize the effort of clinicians and minimize the number and type of annotations required to reach the desired accuracy goal." Enzo delved into the automatic augmentation of dataset annotations and its challenges. Most publicly available medical imaging datasets include only a single type of annotation, like disease labels or anatomical masks. The emergence of vision-language models with open vocabularies presents a promising solution by enabling the generation of diverse annotations, but this raises the question of whether we can fully rely on such approaches. "During the workshop, we explored key aspects of the medical imaging dataset creation process, ranging from technical challenges with multicentric data to the legal considerations that must be addressed," said Enzo, and added: "I led a discussion on the complexities of generating automatic annotations for large-scale databases, with a focus on using NLP for disease tagging from electronic health records and developing anatomical segmentation masks for large scale X-ray databases. In that regard, I shared insights from our recent CheXMask paper, highlighting our experiences in dataset creation and the major technical challenges we encountered, such as implementing automatic quality control and conducting fairness audits without ground truth.“ The workshop also included a virtual session led by Víctor M. Campello (University of Barcelona) and Camila González (Stanford University), who connected with online participants on the first day. "Yes, I recently had the opportunity to moderate - together with Víctor
RkJQdWJsaXNoZXIy NTc3NzU=