Over the past decade, deep learning models have revolutionized image processing by autonomously extracting task-specific features from training data, outperforming traditional machine learning methods that relied on manually crafted features. This shift has significantly advanced medical image analysis. However, most techniques depend on fully supervised training with task-specific labeled datasets, limiting their generalizability across diverse medical imaging modalities, datasets collected from different imaging devices, and task-specific function requirements. Creating high-quality object-level annotations medical datasets requires substantial time and effort, placing a significant burden on radiologists. Medical image analysis often faces the challenge of data scarcity, which limits the applicability of deep learning models to narrow domains.