Artificial intelligence (Al)'s potential applications in medical imaging are vast and significant. One of the chief obstacles to the development and clinical implementation of Al algorithms is the availability of sufficiently large, curated, and representative training data based on expert labeling. Collecting such curated data requires many hours of manual work but is crucial for downstream model performances. Within my PhD, we investigate human-centered machine learning methods enabling physicians to curate high-quality datasets. Our goal is to support physicians in this process by understanding the dataset's limitations and suggesting new samples leading to a more representative and robust dataset. Therefore, we leverage similarity and representation learning methods in an interactive learning approach.