A commentary on “CT-based radiomics signature of visceral adipose tissue for prediction of early recurrence in patients with NMIBC: a multicentre cohort study”

Dear Editor, We read the paper by Yu et al, titled “CT-based radiomics signature of visceral adipose tissue for prediction of early recurrence in patients with NMIBC: a multicentre cohort study”1, with real curiosity. It landed in the International Journal of Surgery (2025). The team built a fresh multimodal prediction model. They mixed radiomic features from visceral adipose tissue (VAT) with the usual clinical numbers. The goal: spot who with non-muscle-invasive bladder cancer (NMIBC) will relapse soon after the first transurethral resection of bladder tumor (TURBT). Their method felt solid. They signed up patients across centers, locked down image settings, and traced every slice the same way. They also ran internal–external validation. That setup makes the numbers travel better. Still, a few bits need a closer look. First, although the study enrolled 325 patients from three tertiary hospitals within a single Chinese province – a relatively large sample size for radiomics research – the cohort exhibits substantial demographic and clinical homogeneity. This geographical and institutional concentration may compromise external validity, as CT acquisition protocols, patient body composition, and tumor biological heterogeneity vary considerably across regions and healthcare settings in China – and globally. Moreover, the retrospective observational design precludes causal inference and introduces the risk of residual confounding. While the authors employed multivariate logistic regression to adjust for established clinical covariates, several biologically plausible confounders – including dietary patterns, physical activity levels, genetic ancestry, specific components of metabolic syndrome (e.g., insulin resistance and dyslipidemia), and prior pharmacologic interventions – were not measured or incorporated. Omission of these variables may inflate the apparent predictive performance of VAT radiomic features, thereby overstating their independent prognostic utility. Second, although intra- and inter-observer agreement for segmentation was robust intraclass correlation coefficient (ICC) > 0.8, radiomic feature stability remains highly sensitive to technical variability – including scanner manufacturer, acquisition parameters (e.g., tube voltage, reconstruction kernel, and slice thickness), and temporal drift across imaging sessions. Critically, the study did not report whether feature harmonization or stability assessment was performed across centers, timepoints, or scanner platforms – nor did it implement standardized normalization or batch-effect correction. Such methodological gaps pose a significant threat to reproducibility and clinical generalizability in multicenter settings. In contrast, recent advances demonstrate that deep learning–based automatic segmentation of subcutaneous and visceral fat compartments achieves high concordance with manual annotations (Pearson product–moment correlation coefficient and ICC > 0.9)2. Integrating such automated, harmonized segmentation into the radiomic workflow would substantially enhance model robustness, cross-institutional portability, and scalability. Third, compelling biological evidence links visceral adiposity to tumor progression via paracrine and systemic inflammatory mechanisms. As highlighted in recent reviews, adipocytes and tumor-infiltrating immune cells within VAT secrete adipokines – including leptin, adiponectin, tumor necrosis factor-α, and IL-1β – that modulate metabolic inflammation, immune surveillance, and stromal remodeling3. Tumor-associated macrophages, in particular, frequently adopt an immunosuppressive phenotype that fosters tumor recurrence and therapeutic resistance4. In urothelial carcinoma, more immune cells pile in, PD-L1 lights up, and the tumor carries heaps of mutations. Patients with that combo often do better, so the tumor immune microenvironment really counts5. The team skipped serum adipokines, inflammatory cytokines, and full tumor immune phenotypes. They never checked immune cell makeup, checkpoint levels, or how cells sit in space. Because of that, the link between VAT radiomic features and recurrence stays flat-out correlative. No one showed the in-between biology. Without molecular, immunological, or functional biomarkers baked in, the study cannot tell if VAT radiomics trace back to fat-driven inflammation, active tumor–stroma crosstalk, or just ghost patterns on a scan. The study has its limits. Still, it pushes VAT radiomics for NMIBC prognosis forward with actually practical care. We used to count how much fat sits around the viscera. This team flips the script. They look at the texture, shape, and scatter of those pixels instead. That shift gives the fat a voice about the patient’s future. They ran the numbers and proved these picture clues beat the usual charts. Nice work, we say. Next round needs shared imaging recipes, hands-off segmentation, and wet-lab proof. Once those boxes are ticked, doctors can swap gut-feel risk for a clean, biology-backed score.