Integrating AI into clinical workflows, particularly in medical image analysis, requires systems that are accurate, reliable, interpretable, data-efficient, and robust across clinical settings. While modern AI models can achieve strong diagnostic performance, their clinical value depends on their ability to generalize across institutions, imaging devices, patient populations, and data-limited medical domains. This project focuses on developing trustworthy and data-efficient AI methods for medical image analysis, with the broader goal of improving diagnostic reliability, workflow efficiency, and patient outcomes. The project will pursue several complementary directions. First, we will develop generative methods for creating images to support self-supervised pretraining, with the goal of improving representation learning when real medical data are scarce, heterogeneous, or difficult to share. Second, we will develop and evaluate explainability methods for medical imaging models, emphasizing clinically meaningful evaluation rather than only technical proxy metrics. Third, we will explore representation learning methods for aligning potentially unpaired data modalities, enabling models to integrate heterogeneous medical information with fewer human-annotated pairings. Finally, we will investigate clinical trajectory prediction methods that model how patient states evolve over time and how multimodal clinical information can support earlier and more reliable prediction of future outcomes. Together, these directions aim to support medical AI systems that are more robust, interpretable, data-efficient, and useful in real-world clinical workflows.