Preview

PDF, English - main document Download (103MB) | Terms of use

Abstract

Introduction: Radiomics aims to extract quantitative features from medical images that capture underlying biological and clinical characteristics. Despite its promise for precision oncology, radiomics research continues to suffer from poor reproducibility and limited generalization across studies, software, and imaging modalities. This thesis addresses these fundamental limitations by systematically analyzing how methodological design choices—such as feature extraction, preprocessing, and model selection—affect the robustness and transferability of radiomic biomarkers. To enable this large-scale methodological investigation, I developed the Radiomics Processing Toolkit (RPTK), a fully automated and open-source framework that standardizes radiomics experimentation and benchmarking across heterogeneous datasets. Using RPTK, I conducted comprehensive evaluations on seven open-source cancer imaging cohorts and demonstrated the framework’s applicability in two clinical studies on lung cancer immunotherapy response prediction and colorectal neoplasia detection. Materials and Methods: My work integrates radiomics analyses performed on retrospective data, including seven public datasets and two proprietary cohorts, comprising 3,189 Computer Tomography (CT) and Magnetic Resonance (MR) scans from 3,116 patients with a total of 3,273 segmented regions of interest (ROI). The seven open-source datasets include retrospective MR and CT cancer imaging data concerning different tasks for cancer classification from 931 patients. The proprietary data collection include a multi-timepoint (prior treatment and during treatment) CT dataset for lung cancer immunotherapy treatment response prediction (the Predict study) consisting of 73 patients and a large-scale CT liver imaging datasets with 1,997 patients investigating in colorectal neoplasia detection (LiverCRC study). The mean patient age was 62 ± 17 years, and 53.6 % of participants were male. Radiomics features were extracted using two independent feature-extraction tools, PyRadiomics and Medical Image Radiomics Processor (MIRP), enabling standardized cross-extractor comparisons in compliance with the Image Biomarker Standardisation Initiative (IBSI). The RPTK framework integrates adaptive preprocessing, standardized feature extraction, and robust feature stability filtering to enhance reproducibility and robustness for subsequent model training. Six machine learning models were trained to predict tumor malignancy, treatment response, colorectal neoplasia, or cancer subtypes based on the selected feature sets from each extractor. The performance of RPTK was tested against a state-of-the-art radiomics tool (AutoRadiomics) and six different deep learning models. Results: Across seven open-source datasets, RPTK outperformed both AutoRadiomics and deep learning models (Residual Networks (ResNet) and Densely connected convolutional Networks (DenseNet)), achieving a mean test Area Under the Receiver Operating Characteristic curve (AUROC) of 0.81 ± 0.12 compared to 0.68 ± 0.15 and 0.60 ± 0.16, respectively. In the Predict study, longitudinal delta-radiomics analysis with RPTK improved early prediction of immunotherapy response compared to single-timepoint analyses with RPTK, and the inclusion of clinical variables further enhanced model performance in RPTK. RPTK achieved a test AUROC of 0.75 ± 0.10 using delta radiomics, outperforming AutoRadiomics (0.51 ± 0.14) and the best deep learning model (0.56 ± 0.14). In the LiverCRC study, RPTK reached a test AUROC of 0.86 ± 0.04, significantly exceeding AutoRadiomics (0.65 ± 0.03) and deep learning (0.60 ± 0.03), demonstrating scalability and generalization in large multi-thousandsample datasets. Beyond these comparisons, RPTK also matched or outperformed 12 additional published test AUROC values reported on the integrated open-source datasets. Conclusion: Collectively, the results demonstrate that RPTK provides robust, state-of-the-art predictive performance across diverse imaging datasets and clinically relevant tasks. Its modular design enables fair cross-framework benchmarking while maintaining flexibility for clinical data integration and ensuring methodological transparency. The open-source release of RPTK fosters community-driven validation and supports future clinical implementation. This work thus represents both a methodological advancement and a step toward reliable, reproducible, and clinically translatable radiomics.