🤖 Generated with AI assistance · Does not replace individualized medical advice.

Deep Learning CT Model Non-Invasively Predicts Tertiary Lymphoid Structures in Pancreatic Cancer

AI Transparency Notice: This post was generated by an AI assistant to summarize and contextualize published medical literature. It is for informational purposes only and does not constitute medical advice, diagnosis, or treatment recommendations. Always consult a qualified healthcare provider for clinical decisions.

Lead

A multicenter study demonstrates that a ResNet50-based deep learning model can accurately predict tertiary lymphoid structure (TLS) status in pancreatic ductal adenocarcinoma (PDAC) using preoperative CT scans. In a prospective immunochemotherapy cohort, the model identified patients with significantly improved objective response rates and overall survival, offering a non-invasive biomarker for immunotherapy selection.

Key Facts

Metric	Value
Model Performance (External Validation)	AUC 0.853 PMID: 42731878
Prospective Pathology Verification	AUC 0.929, Specificity 1.000 PMID: 42731878
OS Benefit (Non-Surgical Cohort)	Median 19 vs 7 months (HR 3.611) PMID: 42731878
ORR Benefit (Prospective Cohort)	90.0% vs 26.5% (p=0.001) PMID: 42731878

Why It Matters / Context

Pancreatic ductal adenocarcinoma remains one of the most lethal malignancies, with limited options for effective systemic therapy. Tertiary lymphoid structures (TLS) are organized aggregates of immune cells within the tumor microenvironment that have emerged as critical drivers of anti-tumor immunity. Historically, the presence of TLS has been associated with improved survival and enhanced responses to immunotherapy. However, detecting TLS requires invasive biopsy and complex histopathological analysis, limiting its utility in routine clinical practice.

The development of a non-invasive, imaging-based biomarker is a significant step toward precision oncology. By leveraging deep learning to extract subtle radiomic features from standard contrast-enhanced CT scans, this study bridges the gap between complex immune microenvironment biology and accessible clinical imaging. The prospective validation in an immunochemotherapy cohort is particularly noteworthy, as it moves beyond retrospective correlation to demonstrate predictive value in a real-world treatment setting. This supports the concept of “radiomic biomarkers” as complementary tools to traditional molecular markers like PD-L1, which often lack specificity in solid tumors.

Details

The study enrolled 223 surgically resected PDAC patients 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 images. The locked model was subsequently evaluated in 44 patients from a prospective immunochemotherapy trial (ChiCTR2000032293).

The model demonstrated robust performance across multiple validation stages. In the training cohort, the area under the curve (AUC) was 0.987, dropping to 0.937 in internal validation and 0.853 in external validation PMID: 42731878. In the non-surgical cohort (n=82), patients predicted as TLS-positive by the DLN had significantly longer overall survival (OS) compared to TLS-negative patients, with a median OS of 19 months versus 7 months (HR=3.611, p<0.0001) PMID: 42731878.

The prospective cohort provided the strongest evidence for clinical utility. Pathology-anchored verification of the DLN predictions yielded an AUC of 0.929 with a specificity of 1.000 PMID: 42731878. Patients with high DLN scores demonstrated superior clinical outcomes compared to those with low scores:

• Objective Response Rate (ORR): 90.0% vs 26.5% (p=0.001) PMID: 42731878.
• Progression-Free Survival (PFS): Median 13.1 vs 5.3 months (p<0.001) PMID: 42731878.
• Overall Survival (OS): Median 17.6 vs 8.7 months (p<0.001) PMID: 42731878.

Multivariable analysis confirmed that the DLN score independently predicted both OS (HR=0.052, p<0.001) and PFS (HR=0.079, p<0.001) PMID: 42731878.

Biologically, the DLN score was orthogonal to PD-L1 expression (rho=0.011, p=0.946), suggesting it captures distinct immune features not reflected by standard PD-L1 assays PMID: 42731878. Instead, the score 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), which are hallmarks of functional TLS architecture PMID: 42731878. It was also inversely correlated with tumor cell density (rho=-0.399, p=0.007) PMID: 42731878.

Cohort	N	Metric	DLN-High/Positive	DLN-Low/Negative	P-Value
Non-Surgical	82	Median OS (months)	19	7	<0.0001
Prospective	44	ORR (%)	90.0	26.5	0.001
Prospective	44	Median PFS (months)	13.1	5.3	<0.001
Prospective	44	Median OS (months)	17.6	8.7	<0.001

Limitations

Critical Appraisal: While the prospective validation is a major strength, the sample size of the prospective immunochemotherapy cohort is small (n=44). This limits the statistical power to detect smaller differences in efficacy and may introduce selection bias. Furthermore, the model was trained on contrast-enhanced CT scans, which may not be universally available or standardized across all imaging centers. The “black box” nature of deep learning models also raises concerns about interpretability; while the study provides biological correlations, the specific radiomic features driving the prediction are not fully transparent. Finally, the study focuses on PDAC; generalizability to other solid tumors or different immunotherapy regimens remains unproven.

Clinical Takeaway

Clinicians should consider that non-invasive CT-based deep learning models may serve as a complementary biomarker to PD-L1 for selecting pancreatic cancer patients likely to benefit from immunochemotherapy, pending further large-scale prospective trials.

Full Reference

Yu H, Li X, Cheng X, Deng Y, Wang Y, Li Y. 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. Journal for Immunotherapy of Cancer. 2026. PMID: 42731878 [DOI: 10.1136/jitc-2026-015705]