This study identified the presence of intratumoural TLSs in patients with GBC as an independent predictor of RFS. This factor has been linked to better prognoses for patients with ICC [12, 20] and has been found to predict improved immunotherapeutic response for ICC and other solid tumours, such as hepatocellular carcinoma and melanoma, regardless of PD-L1 expression and CD8+ T-cell density [12,13,14,15,16]. In this study, three independent clinico-radiological predictors of TLS status (tumour height, liver invasion and arterial-phase hypo-enhancement) were included in the clinical model and eight features were included in the radiomics model. The final combined model integrated these two models and outperformed both when applied separately. Its survival stratification ability in surgical and immunotherapy cohorts was validated.

Choi et al [30] found that liver invasion was associated independently with resection margin positivity in patients with GBC. As per the system of the American Joint Committee on Cancer, gallbladder tumours classified as T3 have grown through the serosa and/or directly into the liver and/or a proximate extra-hepatic structure [31]. In a previous study, liver invasion was associated with the RFS of patients with completely resected GBC in a univariate analysis, but not in a multivariate analysis [32]. Consistent with the results of this study, a larger proportion of TLS-negative than TLS-positive cases showed liver invasion in the present study [33,34,35]. Whereas only two dimensions of tumour size have been considered in previous studies, we evaluated three dimensions and found that tumour height was an independent predictor of the presence of TLSs. On coronal images, tumour height was associated with the length of the adherent gallbladder wall, and thus likely the probability of wall invasion. The height of TLS-positive tumours was less than that of TLS-negative tumours, which may partly explain the more favourable prognoses of the former. Arterial enhancement has not been investigated in GBC, but it has been shown to be associated with ICC outcomes. Min et al [36] found that the ICC mortality and recurrence risks were lower for cases exhibiting diffuse arterial hyper-enhancement on MRI than for those showing peripheral rim enhancement or diffuse hypo-enhancement. We observed diffuse arterial hypo-enhancement in a larger proportion of TLS-negative than TLS-positive tumours, and this feature was associated with poorer prognosis. It is assumed that arterial hypo-enhancement in TLS-negative tumours might reflect less immune infiltration than the TLS-positive tumours, which is associated with a worse prognosis.

To date, no use of an MRI-based radiomics model for GBC survival prediction has been reported; however, CT-based radiomics models have been used for this purpose in several studies. Yin et al [37] established a radiomics model utilising portal venous phase CT images to discriminate benign and malignant gallbladder disease (AUC, 0.81); The model identified two shape features, two grey-level size zone (GLSZM) features, and one grey-level co-occurrence matrix (GLCM) feature as the top five predictive features. Similarly, the radiomics model features in this study were one shape feature, three GLSZM and GLCM features each, and one first-order feature. Meng et al [38] constructed a nomogram to predict the post–surgical resection survival of patients with GBC (AUC, 0.87); their radiomics model included two fissures and one GLSZM and GLCM feature each. Liu et al’s [39] radiomics model used three shape features and one first-order feature from the original images to predict lymph node metastases in GBC prior to surgery. In Gupta et al’s study, medium texture scale parameters, including both mean and kurtosis, or kurtosis alone, may help predict the histological grade and survival of GBC [40]. Thus, radiomics signatures may reflect the heterogeneity, microscopic pathological features, and immunophenotypes of TLS-positive and TLS-negative gallbladder tumours, enabling their differentiation.

In the training cohort, the specificity in the combined model was less than that in the radiomics model, while accuracy and sensitivity were higher. The possible reason may be as follows: The combined model integrates both clinical variables and radiomics features, offering a more comprehensive representation of the data. While this improves overall predictive performance (as reflected in higher accuracy and sensitivity), the inclusion of clinical variables may introduce additional variability, which could slightly lower the specificity. In contrast, the radiomics model is derived solely from imaging features, which may better capture tumour heterogeneity directly related to TLS status, thereby achieving higher specificity. As for the differences in sensitivity and specificity between the training and validation cohorts, it may be explained as follows: In the training cohort, the radiomics model demonstrates better specificity than the clinical model because it leverages high-dimensional features extracted from imaging data, which are optimised to fit the training data. However, in the external validation cohort, the radiomics model’s performance reflects its generalizability, resulting in higher sensitivity but a slight decrease in specificity compared to the training cohort.

This study was retrospective; the MRI dataset from the validation cohort was small, and it was inevitably affected by selection bias, as GBC is a relatively rare disease. However, the combined model was externally validated, demonstrating its reliability. Prospective multicentre studies with larger cohorts are needed. Second, manual segmentation was performed in this study; the reliability and reproducibility of the application of automatic segmentation for liver neoplasms should be explored. Third, the inclusion of clinical variables in the combined model may introduce additional variability, which could slightly lower the specificity. Therefore, the combined model has a limited effect on improving the performance of the radiomics model. Fourth, a difference in sample size in the training and external validation cohorts may cause greater statistical variability in the external validation cohort, affecting the sensitivity and specificity. As a result, a larger and more balanced sample size should be considered in the further study.