This application is a continuation of the current NAISS 2024/6-367. This project is focused on generative AI in different contexts in healthcare. The main challenge of using generative AI in medical imaging lies in ensuring reliability, accuracy, and resilient to missing data and/or modalities. Medical images, such as MRIs or CT scans, are critical for diagnosing diseases, so any AI-generated or AI-enhanced images must be highly precise and free from artefacts that could mislead clinicians. Generative AI systems, however, can sometimes produce images that appear realistic but lack medical validity, creating the risk of false positives or missed diagnoses. Additionally, these systems require extensive, high-quality datasets for training, often paired, which are frequently limited due to privacy concerns and the cost of data curation. Furthermore, ensuring that generative AI models are interpretable, and their outputs are rigorously validated is crucial to safely integrating them into clinical workflows. In this respect, we are investigating how to bridge paired and unpaired image-to-image translation, how to get precise and diverse images while reflecting the textural properties that carry powerful diagnostic and prognostic information, how to mitigate issues given by the sensitivity to different acquisition parameters, which creates data discrepancy in the acquired images which affects the performance of quantitative analysis tools, how to manage 3D medical data generation with an intrinsic volumetric approach exploiting powerful latent space representation empowered by embedding language data information from clinical reports. Applications are directed towards oncology and radiology, where we are working to set up virtual scanning and virtual treatment tools, that translate between imaging modalities and synthesize longitudinal scans for treatment planning.