Quality, Reproducibility, and Value of Artificial Intelligence in Radiology: A Workflow-Oriented Perspective

his doctoral thesis investigates artificial intelligence in radiology from a workflow-oriented perspective, focusing on reproducibility, methodological quality, and clinical value. The work demonstrates that image segmentation is one of the major determinants of radiomics robustness, showing that differences among readers and institutions can significantly affect model performance and generalizability. Strategies based on segmentation revision and observer normalization are proposed to improve the reproducibility of quantitative biomarkers. The thesis also evaluates deep learning–based automated segmentation, demonstrating high performance in muscle quantification and pulmonary disease assessment. In parallel, applications of artificial intelligence for radiological reporting, natural language processing, and large language models are examined. Finally, the methodological quality of AI studies in radiology is critically assessed, leading to the development of evaluation frameworks aimed at improving transparency, reproducibility, and clinical integration.