Interpretable Multi-modal Deep Learning Radiomics for Preoperative Mapping of Tertiary Lymphoid Structure Associated B-Cell Tumor Immune Microenvironment and Immunotherapy Response in Clear Cell Renal Cell Carcinoma

<title>Abstract</title> <p>Background Tertiary lymphoid structures (TLSs) are associated with prognosis and immunotherapy response in clear cell renal cell carcinoma (ccRCC), but their assessment requires invasive tissue sampling. This study aimed to develop and validate a CT-based multi-modal artificial intelligence model for preoperative TLS prediction in ccRCC and to uncover the pathological and molecular determinants of the model’s predictions. Methods This multicenter, retrospective-prospective study enrolled 1,361 patients with ccRCC from eight tertiary hospitals, assigned to primary, test 1, test 2, prospective, multi-omics and immune checkpoint inhibitor (ICI) therapy sets. Arterial-phase CT images were used to train 2D, 2.5D, and 3D deep learning (DL) networks, alongside machine learning (ML) radiomics models. The best single-modality models were integrated via pre-fusion (PFM) and late-fusion (LFM) strategies. Performance was evaluated via the area under the receiver operating characteristic (ROC) curve (AUC). Prognostic value was analyzed via Kaplan-Meier curves. Interpretability was investigated using SHaply Additive Explanations (SHAP) algorithm, HoVer-Net cell segmentation, and bulk RNA sequencing analysis. Results Among the 22 single-modality models, 2.5D-DenseNet161, 2D-ResNet50, Radiomics-XGBoost, and 3D-ResNet101 performed best in test set 1 (AUCs: 0.899, 0.888, 0.878, and 0.756, respectively), which were adopted for fusion models’ establishment. The SVM-based LFM surpassed the PFM (AUC: 0.938 vs. 0.919) and maintained stability across the test 2, prospective, and multi-omics sets (AUC range, 0.877–0.918). LFM-predicted TLS positive significantly correlated with prolonged progression-free survival (all P &lt; 0.05) and effectively discriminated clinical responders from non-responders for ICI therapy (AUC: 0.807). Model predicted TLS positive showed higher inflammatory cell densities (P &lt; 0.05), upregulation of the CXCL13-CXCR5 axis and B-cell-related markers, enrichment of humoral immune response pathways and increased naive and memory B-cell infiltration (P &lt; 0.05). Conclusions The multi-modal LFM provides a non-invasive surrogate for TLS assessment in ccRCC and may support preoperative risk stratification and immunotherapy decision-making.</p>