Development and Validation of a Preoperative MRI Habitat Radiomics Model to Predict Variant Histology in Bladder Cancer.

Abstract

Background: Bladder cancer (BCa) with variant histology (VH) is aggressive, leading to poor prognosis and resistance to neoadjuvant treatment (NAT). Preoperative identification of VH may be important for informing treatment options.

Purpose: To develop and validate a multiparametric MRI-based ensemble model to identify VH in BCa and explore its association with disease-free survival (DFS) and NAT response.

Study type: Retrospective.

Subjects: Six hundred twenty patients with pathologically confirmed BCa (median age, 65 years [IQR: 56, 73], 145 female) from four centers who underwent preoperative MRI, were divided into a training (n = 311), internal validation (n = 54) from Center 1, and three external validation datasets (n = 85, 68, and 102, respectively). Two additional cohorts, DFS (n = 75) and NAT (n = 69) cohorts, were collected from Center 1 to evaluate prognosis.

Field strength/sequence: 3T, non-fat suppressed T2-weighted imaging using fast spin echo, diffusion-weighted imaging using single-shot echo planar imaging, and T1-weighted dynamic contrast-enhanced sequence using 3D gradient echo sequence.

Assessment: Habitat, radiomic, clinical, clinical-radiomic based, and the VHRisk Score (VHRiS) models were constructed for evaluating VH. The prognostic value of VHRiS for DFS and pathological complete response (pCR) rate was further evaluated.

Statistical tests: Mann-Whitney U test, t-test, ROC analysis (AUC), Kaplan-Meier curves, log-rank test, and SHapley Additive exPlanations (SHAP) analysis.

Results: The VHRiS model demonstrated favorable accuracy (AUCs: training, 0.971; internal validation, 0.895; external validation, 0.898-0.974). Low-risk patients (VHRiS ≥ 0.863) exhibited significantly longer DFS than high-risk patients (4.20 months vs. 3.08 months) in the DFS cohort (median follow-up period: 13.19 months [IQR: 6.54, 31.91]). They also showed a higher pCR rate than high-risk patients (64% vs. 33%) in the NAT cohort.

Data conclusions: The VHRiS model may be a robust tool for identifying VH, and may offer a potential method for risk stratification and prognosis prediction in patients with BCa.

Levels of evidence: 4.

Technical efficacy stage: 2.