Deep learning-based CT model for non-invasive prediction of tertiary lymphoid structures in pancreatic cancer: a multicenter study with prospective validation in an immunochemotherapy cohort.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, West China Hospital of Sichuan University, Chengdu, Sichuan, China.
- Health and Management Center, General Practice Medical Center, West China Hospital of Sichuan University, Chengdu, Sichuan, China.
- Division of Abdominal Tumor, Department of Medical Oncology, Cancer Center, Laboratory of Abdominal Tumor Immunology and Microenvironment, State Key Laboratory of Biotherapy, West China Hospital of Sichuan University, Chengdu, Sichuan, China.
- School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, China.
- Sichuan Key Laboratory of Medical Imaging and Department of Radiology, Affiliated Hospital of North Sichuan Medical College, Nanchong, Sichuan, China.
- Functional and Molecular Imaging Key Laboratory of Sichuan Province, West China Hospital of Sichuan University, Chengdu, Sichuan, China.
- Department of Radiology, West China Hospital of Sichuan University, Chengdu, Sichuan, China [email protected] [email protected].
- Division of Abdominal Tumor, Department of Medical Oncology, Cancer Center, Laboratory of Abdominal Tumor Immunology and Microenvironment, State Key Laboratory of Biotherapy, West China Hospital of Sichuan University, Chengdu, Sichuan, China [email protected] [email protected].
Abstract
Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal malignancies. Tertiary lymphoid structures (TLS) have emerged as key components of the tumor immune microenvironment, associated with improved survival and enhanced immunotherapy response. However, non-invasive detection of TLS remains a significant clinical challenge. This multicenter study aimed to develop and validate a deep learning-based CT model to noninvasively predict TLS status and evaluate its prognostic and predictive value across surgical, non-surgical, and prospective immunochemotherapy cohorts. We enrolled 223 surgically resected patients with PDAC from two centers for model development and 82 non-surgical patients for clinical validation. A ResNet50-based deep learning network (DLN) was trained to predict TLS status from preoperative contrast-enhanced CT. The locked model was evaluated in 44 patients from a prospective immunochemotherapy trial (ChiCTR2000032293), with pathology-anchored verification, outcome assessment, and spatial immune profiling via cyclic multiplex tyramide signal amplification. The DLN achieved area under the curves (AUCs) of 0.987, 0.937, and 0.853 in training, internal, and external validation cohorts. In the non-surgical cohort (n=82), DLN-predicted TLS-positive patients had significantly longer overall survival (OS; median 19 vs 7 months; HR=3.611, p<0.0001). In the prospective cohort, pathology-anchored verification yielded an AUC of 0.929 (specificity=1.000). DLN-high patients demonstrated superior objective response rate (90.0% vs 26.5%, p=0.001), progression-free survival (PFS) (median 13.1 vs 5.3 months, p<0.001), and OS (median 17.6 vs 8.7 months, p<0.001). The DLN score independently predicted OS (HR=0.052, p<0.001) and PFS (HR=0.079, p<0.001) in multivariable analysis. The DLN score was orthogonal to programmed death-ligand 1 (PD-L1) expression (rho=0.011, p=0.946) but correlated with dendritic cell-T helper cell spatial interactions (p=0.037) and T helper-B cell spatial co-localization (rho=0.307, p=0.043)-two hallmarks of functional TLS architecture-and inversely with tumor cell density (rho=-0.399, p=0.007). This DLN-based CT model reliably predicts TLS status in PDAC with prospectively validated prognostic and predictive value in immunochemotherapy, biological grounding in TLS-associated spatial immune architecture, and orthogonal predictive value to PD-L1, supporting a complementary biomarker strategy for immunotherapy patient selection.