Radiomics in clinical decision support: unraveling challenges and maximizing benefits for precision medicine

Background: Radiomics, which extracts high-dimensional quantitative data from medical images, holds immense promise for advancing precision medicine. However, its translation into routine clinical practice is significantly hampered by persistent challenges related to the reproducibility and robustness of its workflows. Aims: This thesis systematically addresses these challenges through a threepronged investigation aimed at: 1) enhancing the reproducibility and robustness of radiomics workflows; 2) comparing the performance of traditional machine learning (ML)-based radiomics with modern deep learning (DL)-based approaches; and 3) facilitating the clinical translation of radiomics into practical decision support tools. Methods: The research involved a multi-faceted approach across various clinical domains, including lung, prostate, and breast cancer, as well as multiple sclerosis, utilizing computed tomography (CT), positron emission tomography (PET), and magnetic resonance imaging (MRI). Sources of analytical variability, such as image pre-processing and segmentation methods, were systematically investigated. An automated ensemble modeling pipeline was developed to construct robust predictive models founded on stable features. A head-to-head comparison was conducted between a traditional ML pipeline and a DL architecture for mammographic lesion classification. To promote clinical adoption, a user-friendly, integrated software platform, matRadiomics, was developed. Furthermore, DL models were implemented to automate critical workflow stages like lung and prostate segmentation. Finally, a radiomics-based model was validated as a decision support tool for interpreting uncertain bone uptakes in [18F]PSMA-1007 PET/CT imaging. Results: The initial investigations confirmed that pre-processing parameters and segmentation methodologies are primary sources of feature instability, undermining model reliability. The direct comparison between paradigms demonstrated the markedly superior performance of the DL architecture (average validation AUC > 81%) over traditional ML models (AUC < 68%) for classifying complex lesions. The developed DL automation tools achieved high accuracy, with lung segmentation reaching a Dice score exceeding 0.97, and 0.87 for whole-gland prostate segmentation. Critically, the radiomics decision support tool was shown to significantly improve the diagnostic accuracy of clinicians interpreting PET/CT scans, bringing the performance of less experienced readers closer to that of experts. Conclusion: This thesis provides a comprehensive framework and validated solutions for overcoming key barriers to the clinical implementation of radiomics. By systematically enhancing workflow robustness and strategically leveraging the superior performance of deep learning, this work charts a clear trajectory for transforming radiomics from a promising research concept into a reliable, impactful, and indispensable component of clinical decision support in the era of precision medicine.