| Abstract [eng] |
Medical physicists play an important role in integrating imaging, treatment planning, and therapy monitoring into cancer care. With the increased availability of high-resolution medical imaging and advanced computational tools, radiomics has emerged as a potent non-invasive method for quantifying tumor features. Delta radiomics, which examines changes in imaging features during or after therapy, sheds light on treatment-induced biological impacts. However, the practical application of delta radiomics in ordinary clinical workflows is still limited, primarily due to the complexity of data processing, diversity in radiomic feature extraction procedures, and the lack of defined, clinically validated implementation paths. The aim of this master's thesis was to create and test a reproducible delta radiomics-based process that would assist medical physicists in prognostic modelling for patients with head and neck cancer. Rather than focusing solely on prediction accuracy, the goal was to develop a simple and adaptable approach for clinical usage. In this work, pre- and post-treatment medical imaging data were used to identify delta radiomic features to verify the performance of the developed workflow. Two independent machine learning models were created: one to determine which medication causes the greatest radiomic alterations and another to discover variables linked with patient survival. These modelling tasks not only revealed the workflow's power to recognize clinically meaningful patterns but also emphasized its potential to discover non-invasive imaging biomarkers for treatment monitoring and survival prediction. The developed workflow includes key steps, such as balancing the dataset using the synthetic oversampling method SMOTE, correlation analysis to reduce over-sampling and feature selection using recursive feature removal with Random Forest and XGBoost and evaluating the performance of these methods. The CatBoost method is then used to build classification models based on the given features. Finally, Mann-Whitney U, Kruskal-Wallis and Dunn post hoc tests are used to assess the statistical significance and discriminatory power of the features included in the final models. This thesis advances medical physics by proposing a robust, useable delta radiomics methodology for developing imaging-based prognostic models. It emphasizes the critical role of medical physicists in enabling data-driven, individualized treatment evaluation and response assessment. By increasing the incorporation of quantitative imaging biomarkers into clinical radiation workflows, this study adds to the ongoing enhancement of customized oncology care and the employment of AI techniques in routine clinical practice. |
