Renal pseudocontinuous arterial spin labeling radiomics for early diagnosis and staging of chronic kidney disease

Early chronic kidney disease (CKD) may be difficult to detect with conventional renal function measures, whereas pseudocontinuous arterial spin labeling (pCASL) can noninvasively quantify renal blood flow (RBF) and characterize perfusion heterogeneity. This prospective two-center study aimed to develop and validate pCASL-derived RBF radiomics models for early CKD diagnosis and staging. We included 364 patients with CKD and 41 healthy volunteers (HVs); Center 1 data were split 7:3 into training and internal test sets, while Center 2 data formed an independent external test set. After semi-automated deep-learning and clustering-based segmentation of the renal cortex and medulla with manual refinement, radiomics, clinical, and combined models were developed for HVs versus CKD stage 1, HVs versus CKD stages 1–2, and CKD stages 1–2 versus stages 3–4. The combined model achieved area under the receiver operating characteristic curve (AUC) values of 0.927, 0.754, and 0.929 for HVs versus CKD stage 1; 0.883, 0.901, and 0.775 for HVs versus CKD stages 1–2; and 0.970, 0.952, and 0.984 for CKD stages 1–2 versus stages 3–4 in the training, internal test, and external test sets, respectively. It significantly outperformed individual models in selected analyses but not consistently across datasets, showed satisfactory calibration in the training and internal test sets, and suggested potential net clinical benefit in exploratory training-set analyses. Combining pCASL-derived RBF radiomics with clinical variables may support noninvasive early CKD assessment and staging; however, the small external cohort, particularly for stage 1 disease, necessitates validation in larger, more diverse multicenter populations.