Verdicchio, Mario (2025) Artificial Intelligence in Cross-Modal Biomedical Imaging: Interpretability, Data Efficiency, and Modality Synthesis. [Tesi di dottorato]

Documento PDF Verdicchio_Tesi_PhD_firmata.pdf Visibile a [TBR] Amministratori dell'archivio Download (56MB) | Richiedi una copia

Abstract

Artificial Intelligence (AI) is progressively revolutionizing the medical imaging sector. Translating AI into concrete clinical applications requires a deep and multidimensional investigation, encompassing technological, methodological, and organizational aspects. This thesis is the result of a training and research path focused on three specific aspects: model interpretability, data management efficiency to ensure reliability and scalability, and the application of generative models for the synthesis of unavailable imaging modalities. In the first domain, the thesis addresses the issue of interpretability through the development of a “pathomic” framework for prostate carcinoma grading. This approach integrates multi-scale feature extraction with explainability techniques, with the goal of building a diagnostic model whose decision logic is transparent, inspectable, and certifiable. Regarding data efficiency, a new and efficient training data selection strategy is proposed and validated. It allows for the identification of a minimal yet highly informative data subset, reducing the annotation burden required by deep learning models. Finally, to address the challenge of multimodal synthesis, a generative AI framework has been developed and applied to the neuroradiological domain. The model is capable of generating Magnetic Resonance Imaging (MRI) scans from Computed Tomography (CT) images, enabling quantitative assessments, such as cerebral atrophy estimation, in clinical settings lacking MRI scanners. Overall, this work explores how AI applications can contribute to the evolution of biomedical imaging.

Downloads

Actions (login required)

Modifica documento