Preoperative CT radiomics–deep learning signature for lymphovascular invasion in bladder cancer: External validation and clinical utility.

Abstract

857 Background: Lymphovascular invasion (LVI) critically guides neoadjuvant therapy, cystectomy timing, and surveillance in bladder cancer, yet it is rarely known preoperatively. We developed and externally validated lymphovascular invasion radiomics–deep learning signatures (LRDs) to noninvasively predict LVI and provide a clinicopathologic+LRDs nomogram. Methods: Retrospective two-center cohort of 302 bladder cancer patients with preoperative contrast-enhanced CT and pathologic LVI reference. QLYY split 7:3 (train/internal); TCIA external. CTs standardized; radiomics+DL features stacked. Performance (AUC, calibration, DCA) and survival (KM/Cox) evaluated. Results: 302 patients (training/internal/external 174/75/53; male 76%/87%/81%). Baseline features were comparable; the external cohort was older (69.3 y; SMD_age≈0.30) with higher LVI (54.7% vs 41–43%), more MIBC (54.7% vs 37–41%) and N+ disease (34% vs 19–25%), forming a tougher validation set. Among single-source models, radiomics-MLP outperformed DL-only (AUC 0.856/0.859/0.832 train/internal/external vs 0.758/0.746/0.773 for DL-LR). The stacked LRDs improved discrimination to 0.859 (95%CI 0.806–0.913) in training, 0.872 (0.792–0.952) internally, and 0.872 (0.773–0.972) externally, with ΔAUC vs radiomics +0.003/+0.013/+0.040 and vs DL +0.101/+0.126/+0.099. DeLong favored LRDs over DL and showed gains over radiomics (largest externally). DCA showed highest net benefit at pt = 0.20–0.60 with near-ideal calibration; results were stable in NMIBC/MIBC. Clinically: ≤0.20 rule-out (bladder-sparing), ≥0.60 rule-in (early RC/NAC). NRI confirmed incremental value (LRDs vs DL +0.159/+0.200/+0.075; vs radiomics +0.015/+0.044/+0.055 for train/internal/external). A clinicopathologic+LRDs nomogram achieved AUC 0.912 (95%CI 0.871–0.954) in training and 0.886 (0.808–0.965) internally. In multivariable analysis, MIBC, nodal positivity, HER2 overexpression, and LRDs (per-SD) were independently associated with LVI (all p < 0.05). LRDs generated clinically meaningful probabilities (e.g., 0.28 in pathologic LVI−, 0.65 in LVI+), aligning with the rule-out/rule-in cut points above. Prognostically, high LRDs predicted worse outcomes (PFS HR 2.35, 95%CI 1.55–3.57; OS HR 3.10, 1.90–5.06; both p < 0.001), with time-dependent AUCs 0.698/0.718/0.724 (PFS) and 0.730/0.741/0.781 (OS) at 1/3/5 years. Conclusions: The integrated CT-based LRDs accurately estimates preoperative LVI, is externally validated in a harder cohort, and adds independent value beyond pT/pN/HER2. With strong calibration and higher net benefit than single-source models, and deployable from routine CT via a clinicopathology+LRDs nomogram, it enables individualized counseling, guides perioperative planning, and stratifies PFS/OS risk.