Abstract
Imaging is increasingly essential for the diagnosis, prognostication, and management of cancer patients. It is a field in rapid evolution and has the potential to provide noninvasive characterization of malignancy, which remains an important Holy Grail of medicine. Radiomics, which involves the high-throughput assessment of imaging features, aims to transform images into structured phenotypic data. By correlating imaging-derived data with clinical and molecular information, radiomics offers the potential to relate tumor biology to imaging phenotype to improve patient prognostication and management. This review surveys the principles and applications of classical and deep learning–based radiomic approaches in oncology, and it highlights key challenges in robustness and reproducibility that must be addressed to realize the potential of radiomics in clinical practice.