An MRI-based model for preoperative prediction of tertiary lymphoid structures in patients with gallbladder cancer.
Ying Xu, Zhuo Li, Weihua Zhi, Yi Yang, Jingzhong Ouyang, Yanzhao Zhou, Zeliang Ma, Sicong Wang, Lizhi Xie, Jianming Ying, Jinxue Zhou, Xinming Zhao, Feng Ye
Insights into Imaging · August 30, 2025 · Vol 16 · Issue 1 · p. 189
Clinical question
Can preoperative MRI-based radiomics predict intratumoural TLS status in GBC, and can such a model stratify recurrence-free survival (RFS) after surgery and overall survival (OS) during immunotherapy?
Population
Training cohort: 129 GBC patients from centre 1 (December 2012–November 2022); external validation cohort: 44 GBC patients from centre 2 (February 2015–December 2021); immunotherapy cohort: 12 advanced GBC patients from a phase-Ib clinical trial (NCT03825705). Median age 62 years; ~45% male in both main cohorts.
Methods
Radiomics features were extracted from six MRI sequences (non-enhanced T1W, T2W, DWI, arterial, portal venous, and delayed phases) via manual VOI segmentation. After variance and correlation filtering (642→78 features), multivariate logistic regression selected 8 features for the Rad-score. Clinical predictors were identified by univariate/multivariate logistic regression. A combined model integrated Rad-score with three clinico-radiological variables. Model performance was assessed by AUC, calibration curves, Hosmer–Lemeshow test, and DCA; survival was analysed by Kaplan–Meier and log-rank test.
Key findings
Intratumoural TLS presence was the only independent predictor of RFS in multivariate Cox regression (HR 1.86; 95% CI 1.01–3.43; p=0.046); TLS-positive patients had significantly longer RFS than TLS-negative patients (p=0.002).
Evidence grounded · 2 source references
Three independent clinico-radiological predictors of TLS status were identified: tumour height (OR 0.67), liver invasion (OR 0.37), and arterial-phase hypo-enhancement (OR 0.33). Eight radiomics features formed the Rad-score; TLS-positive tumours had significantly higher Rad-scores than TLS-negative tumours in both cohorts (p<0.001).
Evidence grounded · 2 source references
The combined model (AUC 0.891 in training) outperformed the clinical model (AUC 0.870) and radiomics model (AUC 0.775) in the training cohort and was externally validated. Low-risk patients had median RFS of 59.6 months vs 8.7 months for high-risk in the training cohort, and 18.4 vs 4.2 months in the external validation cohort.
Evidence grounded · 3 source references
In the immunotherapy cohort, low-risk patients had a significantly longer median OS than high-risk patients (24.7 months vs 10.5 months), demonstrating the model's utility for predicting immunotherapy outcomes.
Evidence grounded · 2 source references
Limitations
Retrospective design with small external validation cohort (n=44) and small immunotherapy cohort (n=12), susceptible to selection bias; manual segmentation limits reproducibility; inclusion of clinical variables may slightly reduce specificity; imbalanced cohort sizes increase statistical variability.
Evidence grounded · 1 source reference
Conclusion
An MRI-based combined radiomics-clinical model can preoperatively predict intratumoural TLS status in GBC, outperforming individual models, and accurately stratifies postoperative RFS and post-immunotherapy OS, serving as a non-invasive preoperative TLS marker.
Radiology relevance
Multiparametric MRI (six sequences including DWI and multiphase contrast-enhanced sequences) enables non-invasive preoperative assessment of tumour immune microenvironment features (TLS status) in GBC, replacing the need for postoperative histopathological TLS evaluation and informing surgical and immunotherapy treatment decisions.
Evidence grounded · 3 source references
Teaching points
Arterial-phase hypo-enhancement on MRI is an independent predictor of TLS-negative GBC and is associated with poorer prognosis, possibly reflecting reduced immune infiltration; this MRI feature, combined with tumour height and liver invasion, can help identify patients at higher recurrence risk without tissue sampling.
Evidence grounded · 2 source references