Towards explainable radiomics: stability and interpretability in computational models for clinical applications

Radiomics has been widely used in medical imaging for various tasks, like biomarker prediction, tumor subtype classification, and forecast of disease recurrence and treatment response. However, increasing literature shows that the radiomics approach is rarely explored from a methodological perspective. This thesis focuses on key challenges of radiomic analysis, encompassing segmentation, evaluation of feature stability, construction of explicable prediction models, and image harmonization, with the main scope of investigating radiomics reliability and contributing to the development of interpretable radiomics-based models for prediction purposes in clinical applications. Radiomics reliability is addressed in terms of stability with respect to segmentation variability and stability of the predictive performance. The proposed machine learning models focus on the explainability of feature selection for the prediction of triple-negative breast cancer subtype, and on the interpretability of model architecture for radiomics-based longitudinal prediction of glioblastoma response to treatment. Other contributions concern numerical schemes for medical image segmentation and deep learning Image2Image network for image harmonization, both conceived as a preliminary step to radiomic analysis. While the investigation conducted in this thesis deals with specific applications, the methodology introduced and the findings reported could potentially provide a broader understanding of the principles underlying radiomics-based analyses. Ultimately, this work aspires to contribute to a deeper comprehension of the interplay between feature stability and segmentation variability for model explainability. Moreover, the construction of interpretable models and harmonization strategies represents a step toward refining the radiomics workflow, fostering the development of more reliable, generalizable, and clinically applicable computational models for diagnosis and prognosis.