Abstract
dc:descriptionThe integration of AI and radiomics opened new horizons within the current paradigm of managing diagnosis and prognosis in lung cancer. In such a context, the present doctoral work undertakes serious problems in the sector of precision oncology. It tends to propose innovative for the early modeling and clinical interpretation of treatment response in patients who underwent radiation and assessment of risks developing of radiation-induced pneumonitis (RP) in non-small-cell-lung cancer (NSCLC) patients. This research initially presents an enormous opportunity using radiomics in uncovering hidden quantitative biomarkers within medical images that enhance the characterization of NSCLC, thus allowing a forecast of outcomes pertaining to treatment response. Using an advanced AI deep learning approach, the researchers show remarkable enhancement of RP forecasting with an AUC-ROC at 0.87, an improvement over past machine learning analysis. Based on these successes, this thesis goes a step further to develop the first framework that marries multi-modal radiomic data-specifically, PET/CT imaging-with interpretable AI methods, including SHAP and LIME. The proposed hybrid model approach not only enhances the analytical performance of AI models in their prognostic capacity and outcome stratification but also provides far-reaching insights into the decision-making process toward ensuring clinical trust and collaboration. These models achieved a Multi-Modal AUC-ROC of 0.90, further establishing their efficiency in distinguishing pathological variations with interpretability. To support the development and validation of the proposed AI-based radiomic models, this doctoral work employed a combination of publicly available and private clinical imaging datasets. Public data sources included repositories such as The Cancer Imaging Archive (TCIA), with NSCLC-Radiomics and RIDER datasets being central to initial experimentation. In addition, private clinical repositories were accessed through established collaborations between our research team at the Medical Physics Department (MPD) at the University Hospital of Larissa, Greece. The MPD is actively engaged in clinical practice, research, and education, and collaborates with the private healthcare sector to acquire imaging datasets of patients diagnosed with lung cancer and radiation-induced pneumonitis. These clinical datasets include a wide range of CT scans and radiographic studies from individuals undergoing chest radiotherapy for various thoracic malignancies. All patient data were handled in strict accordance with data protection and ethical guidelines, ensuring full anonymization, secure storage, and encrypted handling procedures. The entire data lifecycle adhered to the requirements of the General Data Protection Regulation (GDPR), thereby safeguarding patient confidentiality while supporting robust and ethically grounded scientific exploration. This thesis, for the first time, proposes GANs and self-supervised learning to generate synthetic data that will be helpful in conducting robust model training with limited annotated datasets, thus overcoming inherent challenges of data scarcity. These approaches were found instrumental for enhancing detection and localization of RP-affected regions with very high accuracy, setting new benchmarks for early diagnosis and personalized treatment planning in radiation oncology. The study describes a paradigm shift in the translational use of AI-powered radiomics by laying out a holistic framework that bridges the gap between prognostic insight and clinical applicability. Thus, these results open new opportunities for academic and clinical fields on a path toward truly precision oncology, non-invasiveness, transparency, robustness, and clinical value that has not existed before.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Raptis, Sotirios
- Ράπτης, Σωτήριος
Subjects
dc:subject × 11Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59275