Bridging the Gap in Radiomics: A Dual Approach to Reproducible Research and Open-Access Data in Cancer Imaging

Artificial Intelligence methods are revolutionizing the field of radiology. Quantitative image analysis with radiomics features is a particularly promising area of research, as evidenced by numerous significant publications. For successful clinical translation, the models need to be developed and evaluated carefully with robust statistical methods using diverse, properly curated datasets. Our prior research experience highlighted a striking lack of comprehensive tools combining standard radiomics analysis with state-of-the-art machine learning tools. The first part of this dissertation addresses this problem by detailing the process of developing and validating an open-source toolkit for reproducible radiomics research. We developed a comprehensive Python code library that adheres to the best practices for validating research hypotheses. The library covers all the functionalities of the radiomics workflow, including image preprocessing, segmentation, feature extraction, feature selection, modeling, and evaluation. It automatically selects the optimal model parameters for a specified task. Additionally, it features a user-friendly web application that enables a complete radiomics analysis without necessitating any programming experience. This research toolkit was validated carefully across seven clinical applications using eight public datasets, successfully replicating the previously reported results with a fully automated setup. The second part of this dissertation complements the former with a comprehensive review and creation of a public repository with ready-to-use radiomics data. To this end, 143 public cross-section cancer imaging datasets were reviewed, and 29 were selected as suitable for radiomics analysis. Each dataset was subjected to meticulous preprocessing and radiomics feature extraction using AutoRadiomics with a few sets of extraction parameters. An informative wiki with metadata for each dataset was curated to facilitate accessibility and collaboration. The main result is a reference registry with radiomics features and specific preprocessing recipes for radiomics datasets, increasing the utility and visibility of the publicly available cancer imaging datasets. In conclusion, the results of this thesis lay a foundation for advancing the quality, accessibility, and scope of radiomics research. Building upon it, future studies will facilitate the integration of robust radiomics models into clinical oncology practice.