AI in Medical Imaging: From Image Synthesis in MRI to Semantic Analysis in PET-CT

Artificial intelligence (AI) is transforming modern medical imaging by enhancing diagnostic accuracy, improving image quality, and supporting clinical decision-making. This thesis contributes to this transformation through the investigation and development of AI methodologies for Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET), with the strategic objective of supporting and extending the clinical value of contrast agents routinely administered to enhance the visibility of anatomical structures or abnormalities. The first part of the thesis focuses on MRI and introduces novel AI-based approaches aimed at enhancing or simulating the effects of contrast agents. In this context, contrast amplification models (“contrast boosters”) were developed to selectively improve the visibility of brain lesions in diagnostic MR images in both adult and pediatric patients. Additionally, an approach termed “Virtual Contrast” was investigated to assess the ability of AI algorithms to simulate the effects of contrast agent administration using only non-contrast images. These studies demonstrate that AI can significantly augment the effects of contrast agents, with potential improvements in sensitivity and diagnostic power. However, the results indicate that Virtual Contrast methods may have limited generalizability across different tumor types and patient populations, suggesting that contrast agents continue to play an important and currently irreplaceable role in clinical practice. The second part of the thesis addresses PET-CT and presents a framework for quantitative image interpretation. A statistical detection approach was developed that integrates automated organ segmentation with lesion detection to enable reproducible and quantitative analysis. The methodology was validated on heterogeneous datasets encompassing multiple pathologies and tracers, in particular FDG and PSMA, demonstrating robust generalizability. In addition, this work provides insights into the clinical reasoning and domain expertise required for PET-CT interpretation, as well as for the design, development, and evaluation of quantitative imaging software. The thesis delivers two complementary AI applications, generative MRI contrast enhancement and PET CT lesion quantification, and provides both methodological advances and clinical insights. Overall, this work demon- 3 strates how AI-driven methods and contrast agent technologies can be jointly leveraged to improve image interpretation and quantitative analysis in medical imaging.