AN OPERATING SYSTEM FOR AI-BASED APPLICATIONS IN RADIOLOGICAL AND ONCOLOGICAL IMAGING

This doctoral thesis presents an integrated and multidisciplinary research path focused on the development, customization, and application of advanced computational tools for radiological workflows, with particular emphasis on open-source software environments and the integration of artificial intelligence. At the core of this work lies the design and implementation of DicomOS, a Linux-based operating system specifically conceived to meet the practical and technological needs of modern radiology departments. DicomOScombines a stable and customizable infrastructure with a suite of graphical and command-line applications, enabling clinicians and researchers to perform tasks such as annotation, anonymization, DICOM-to-image conversion, dataset organization, and longitudinal analysis with greater e!ciency and transparency.The system was designed with the dual aim of improving clinical workflow and supporting research activities. Its architecture facilitates the integration of existing tools and allows the development of new ones through modular interfaces. Particular attention was dedicated to the graphical user interfaces, which were conceived to lower the technical barrier for non-expert users, while maintaining advanced functionality for experienced researchers. The platform also supports command-line interaction, offering full controlfor scripting and automation in large-scale image processing pipelines. The second major contribution of this work lies in the theoretical and practical exploration of artificial intelligence methodologies in the context of radiological image analysis.The thesis offers an in-depth review of both shallow and deep learning classifiers, discussing their respective advantages, limitations, and appropriate application contexts. Asignificant focus was placed on explainable artificial intelligence, which plays a crucial role in increasing trust, safety, and interpretability in clinical applications. Techniques such as SHAP, Grad-CAM, and attention-based methods were critically examined and applied to real-world medical scenarios.Subsequently, the research addressed the training of Vision Transformer models in constrained data environments. By integrating traditional and diffusion-based data augmentation techniques, robust performance improvements were achieved in the classification of melanoma and mammographic images. These approaches were validated through multiple studies and conference contributions, highlighting the potential of ViT architectures even in low-data regimes typical of medical imaging.In the final part of the thesis, the focus shifts toward the use of graph-based representations of brain connectivity derived from diffusion-weighted imaging. By extractingstructural connectomes and applying both classical machine learning algorithms and graph neural networks, the thesis proposes an original framework for the classification of patientswith traumatic brain injury. The analyses consider not only inter-group discrimination (acute, chronic, controls), but also longitudinal recovery trajectories. To ensure transparency and clinical relevance, explainability methods were employed to localize the most influential brain regions contributing to model predictions, offering new insights into theneural substrates of consciousness and recovery. Taken as a whole, the contributions presented in this thesis advance the field of computational radiology by offering a complete, modular, and explainable framework that bridges the gap between system-level innovation and artificial intelligence methodologies.The proposed tools and methods promote reproducibility, adaptability, and interdisciplinary collaboration, supporting both clinical practice and scientific discovery.