Radiomics-Driven Analysis and Phantom Design for Lung MRI and Diffusion Imaging

Magnetic Resonance Imaging (MRI) is a core biomedical imaging modality that provides high soft-tissue contrast and multiparametric information without the use of ionizing radiation. The increasing complexity and dimensionality of MRI data have progressively driven a shift from qualitative image interpretation toward quantitative and data-driven approaches, with radiomics emerging as a key methodology for imaging-based biomarker development. However, radiomic features are often sensitive to acquisition parameters and methodological choices, raising concerns regarding reproducibility, interpretability, and clinical translation, particularly in the context of advanced MRI-derived metrics and anatomically complex regions. This thesis addresses key challenges in MRI-based imaging through a series of methodological and validation studies spanning controlled phantom experiments and clinically relevant applications. First, an anatomically informed MRI lung phantom was developed to reproduce thoracic tissue magnetic properties relevant to lung cancer imaging and radiomic analysis. The phantom achieved physiologically reliable T1 and T2 relaxation values, demonstrated temporal stability and cross-scanner reproducibility, and provided a controlled platform for protocol optimization and multi-site validation. Second, the feasibility of non-invasive prediction of PD-L1 expression in non-small cell lung cancer was investigated using radiomic and deep learning features extracted from IVIM parametric maps and T1-VIBE anatomical acquisitions. A statistically robust pipeline tailored to small datasets was implemented. Fusion models combining radiomic and deep learning features achieved the best performance; moreover, AUC values ranged from 0.72 to 0.92 across the investigated scenarios, highlighting the potential of MRI-based biomarkers for immunotherapy stratification. Finally, a proof-of-concept investigation into the physical interpretability of radiomic features derived from advanced diffusion MRI was conducted using a liquid crystal phantom with well-defined and evolving diffusion properties. The results demonstrated that radiomics can capture diffusion-related changes beyond conventional scalar metrics, while a dedicated statistical framework enabled the separation of genuine physical effects from acquisition- and inversion-related artifacts. Overall, this work demonstrates how MRI and radiomics can be integrated across different levels of complexity to support the development of more robust and interpretable imaging biomarkers, while clarifying the strengths and limitations of radiomics in modern MRI research.