Radiomics is not dead: quantitative imaging in the era of deep learning in renal cell carcinoma

A growing consensus in the imaging community holds that radiomics has been rendered obsolete by foundation models and end-to-end deep learning. We disagree, and this perspective explains why. The critics are partially right: inter-operator segmentation variability has compromised reproducibility, methodological quality has been inconsistent, and the gap between published AUCs and clinical implementation has been wide. However, they draw the wrong conclusion. In renal cell carcinoma (RCC), where preoperative contrast-enhanced CT is already standard of care for every patient undergoing nephrectomy and where approximately 30% will recur after surgery, radiomics remain an immediately actionable, interpretable, and deployable non-invasive prognostic tool available. We argue that its primary technical limitation (segmentation variability) is being eliminated by the same deep learning technology used to argue against it: automated segmentation frameworks such as nnU-Net achieve Dice coefficients exceeding 0.93 on external cohorts, removing inter-operator noise. We further argue that in low-prevalence, high-stakes oncological settings, the data volumes required to pre-train foundation models do not exist, and that interpretability — now a regulatory requirement under the EU AI Act for high-risk clinical AI systems — is native to handcrafted radiomics and must be retrofitted onto deep learning architectures. Most critically, we present biological evidence that CT images encode tumor microenvironmental and molecular information invisible to the human eye. This generates a directly testable hypothesis: that preoperative radiomic scores can predict not only recurrence but also likelihood of benefit from adjuvant treatment. Radiomics is not dead. It is, at last, being done properly.