Artificial intelligence and immune topography transforming prognostication in pancreatic cancer

Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal human malignancies, characterized by late diagnosis, intrinsic therapeutic resistance, and an extraordinarily complex tumor microenvironment (TME) (1). Despite steady advances in systemic therapies and surgical techniques, the 5-year survival rate remains below 12%. The limited effectiveness of cytotoxic and targeted therapies in PDAC has redirected attention toward the immune contexture of the disease—specifically, the spatial architecture and density of tumor-infiltrating lymphocytes (TILs) within the desmoplastic stroma. Yet, quantifying this immune landscape has long been hampered by technical variability, interobserver differences, and the sheer complexity of histologic interpretation (2).

In this context, the recent study by Kim et al. in JAMA Surgery represents an important step toward merging computational pathology with tumor immunobiology (3).

Artificial intelligence (AI)-based analysis of routine hematoxylin and eosin (H&E)-stained sections reveals distinct immune topographies within pancreatic tumors, including immune-inflamed phenotype (IIP), immune-excluded phenotype (IEP), and immune-desert phenotype (IDP). By capturing the spatial localization of TILs relative to tumor and stromal compartments, AI-derived immune architecture complements and in some cases may supersede traditional anatomy-based tumor-node-metastasis (TNM) staging for prognostic stratification (Figure 1).

Figure 1 AI-driven spatial immune mapping redefining prognostication in pancreatic ductal adenocarcinoma. AI, artificial intelligence; TNM, tumor-node-metastasis.

Using an AI-powered spatial analysis system (Lunit SCOPE IO), the authors analyzed H&E-stained whole-slide images (WSIs) from 304 patients who underwent R0 resection for PDAC at Samsung Medical Center (Table 1). The AI algorithm automatically segmented tumor and stromal compartments, identified TILs, and classified each tumor into one of three immune phenotypes (IPs): IIP, IEP, and IDP. The prognostic relevance of these phenotypes was then assessed with respect to overall survival (OS) and recurrence-free survival (RFS).

The findings of Kim et al. are both clinically meaningful and conceptually provocative. Among all patients, 85.2% of tumors displayed an immune-excluded pattern, with lymphocytes confined to the peritumoral stroma, while only 9.9% were immune-inflamed and 4.9% immune-desert. These proportions are consistent with the well-recognized “immunological coldness” of PDAC, where dense stromal barriers, hypovascularity, and immunosuppressive cell populations prevent effective lymphocytic infiltration (7).

The IIP was associated with dramatically improved survival outcomes: median OS and RFS were not reached during follow-up, in contrast to median OS of 35.1 months for IEP and 11.6 months for IDP (P<0.001). High intratumoral TIL density, but not stromal density, was independently correlated with superior OS and RFS, suggesting that it is not the mere presence but the spatial localization of lymphocytes that determines prognosis. The AI-generated phenotype remained a significant predictor of outcome after multivariable adjustment, with hazard ratios exceeding 5.0 for IDP compared with IIP (8).

By combining pathologic stage and immune phenotype, the authors further demonstrated that stage II tumors with IIP outperformed stage I tumors lacking IIP, implying that immune architecture can modify, or even override, conventional TNM staging. Such an observation challenges the traditional anatomic paradigm and argues for incorporating spatial immune features into future staging or risk-stratification systems (9).

Traditional TIL assessment is laborious, semi-quantitative, and susceptible to interobserver variation. AI-based analysis provides objectivity and scalability, particularly for cancers such as PDAC, where TIL densities are low and manual counting is impractical. By operating directly on routine H&E slides, the approach avoids additional cost and obviates the need for immunohistochemistry (10). The reproducibility and speed of such algorithms can democratize immune quantification, allowing widespread clinical deployment even in high-throughput pathology laboratories (11).

Moreover, the algorithmic classification used by Kim et al. correlates closely with established immunophenotypic frameworks. It includes the tumor-agnostic immune score (IS) system used in non-small cell lung cancer and colon cancer, indicating a potential for cross-tumor generalizability (12). The study thus bridges a crucial translational gap to transform histopathologic morphology into digital, quantitative biomarkers with real prognostic power.

Nonetheless, while the AI’s performance was robust, it remains limited by its reliance on morphological proxies for immune function. The algorithm cannot yet distinguish between CD8⁺, CD4⁺, regulatory, or exhausted T-cell subsets, each of which may carry distinct prognostic or therapeutic implications (13). Integration of multimodal data (e.g., multiplex immunofluorescence or spatial transcriptomics) with AI-derived morphologic analysis may provide a more comprehensive and mechanistic understanding of the TME.

The biological logic underlying the observed prognostic gradient from inflamed to excluded to desert phenotypes was consistent with immune-oncologic principles. The immune-inflamed tumors likely represent those with preexisting T-cell infiltration and cytolytic activity, supported by the authors’ analysis of The Cancer Genome Atlas (TCGA-PAAD) dataset showing elevated CD8⁺ fractions, higher cytolytic activity, and interferon-γ expression in IIP cases (14). In contrast, immune-excluded tumors exhibit stromal sequestration of immune cells, reflecting mechanical or biochemical barriers such as dense fibrosis, tumor-associated fibroblasts, or aberrant extracellular matrix deposition. Immune-desert tumors lack both infiltrate and signaling, representing the most immunologically inert subset (15).

From a clinical standpoint, the study suggests several implications. First, AI-derived immune phenotyping could refine postoperative prognostic stratification beyond TNM staging, identifying patients at higher recurrence risk despite early-stage disease. Second, the differential response to adjuvant therapy observed across phenotypes, with chemotherapy benefiting IIP and IEP groups but not IDP, implies that immune architecture may inform treatment selection. Third, if validated on preoperative biopsy samples, such technology could aid neoadjuvant therapy planning and potentially predict response to immune checkpoint inhibitors (ICIs), which remain largely ineffective in unselected PDAC populations.

Strategies under investigation include stromal remodeling, CXCR4 or TGF-β pathway inhibition, and modulation of myeloid-derived suppressor cells to enhance T-cell infiltration. AI-powered immune phenotyping may serve as a valuable biomarker to monitor these interventions dynamically and noninvasively (16).

Kim et al. deliver a technically rigorous and statistically robust analysis, incorporating an adequately sized, homogeneous cohort (R0-resected PDAC, n=304), long-term follow-up (median 35 months), and comprehensive sensitivity testing. The study adheres to STROBE reporting standards and incorporates external validation through TCGA transcriptomic data, lending credibility to its biological interpretations. Importantly, the work demonstrates practical clinical feasibility, using H&E slides available in every surgical pathology archive—a critical step toward real-world implementation.

Equally noteworthy was the demonstration that AI can transform “routine pathology” into quantitative immunology, expanding the role of digital pathology beyond diagnostic assistance into prognostic and therapeutic prediction. As AI models gain regulatory approval for cancer diagnostics, this study exemplifies how machine learning can augment pathologists by providing reproducible, interpretable metrics derived from massive pixel-level datasets.

The integration of AI-based immune phenotyping into clinical workflow raises several operational and ethical considerations. Implementation requires digital slide scanners, standardized staining protocols, and robust data governance frameworks. However, the scalability of automated analysis and the ubiquity of H&E slides make this approach inherently accessible once infrastructure is in place (17).

From a clinical research perspective, AI-derived immune phenotypes could serve as digital endpoints in prospective trials, allowing automated and unbiased measurement of TME dynamics during therapy. Moreover, harmonizing AI models across institutions via federated learning could accelerate multicenter validation while preserving patient privacy (18).

Looking forward, the convergence of AI-driven morphology, genomic annotation, and spatial transcriptomics will likely yield hybrid biomarkers that capture both structure and function of the TME. In PDAC, where immune exclusion remains a dominant barrier to immunotherapy, such integration may illuminate new therapeutic vulnerabilities.

The study by Kim et al. marks a pivotal moment in surgical oncology, where AI meets tumor immunology to address one of medicine’s most intractable malignancies. By automating the spatial analysis of immune phenotypes, the authors have demonstrated a feasible, scalable, and prognostically powerful biomarker derived from the simplest and most ubiquitous clinical material—the H&E slide. The promise of this approach lies not only in prognostication but also in enabling a deeper mechanistic understanding of the immune microenvironment, guiding therapy, and ultimately, improving patient outcomes. The challenge now lies in translating digital precision into therapeutic precision—validating these insights across diverse populations and integrating them into the next generation of clinical trials and treatment algorithms.

Acknowledgments

None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Gland Surgery. The article has undergone external peer review.

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-511/prf

Funding: None.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-511/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Sarantis P, Koustas E, Papadimitropoulou A, et al. Pancreatic ductal adenocarcinoma: Treatment hurdles, tumor microenvironment and immunotherapy. World J Gastrointest Oncol 2020;12:173-81. [Crossref] [PubMed]
2. Li B, Zhang Q, Castaneda C, et al. Targeted Therapies in Pancreatic Cancer: A New Era of Precision Medicine. Biomedicines 2024;12:1687. [Crossref] [PubMed]
3. Kim H, Choi JH, Lim Y, et al. Artificial Intelligence-Powered Spatial Analysis of Immune Phenotypes in Resected Pancreatic Cancer. JAMA Surg 2025;160:884-92. [Crossref] [PubMed]
4. Truong LH, Pauklin S. Pancreatic Cancer Microenvironment and Cellular Composition: Current Understandings and Therapeutic Approaches. Cancers (Basel) 2021;13:5028. [Crossref] [PubMed]
5. Zhou C, Li J, Wu Y, et al. High Density of Intratumor CD45RO⁺ Memory Tumor-Infiltrating Lymphocytes Predicts Favorable Prognosis in Patients With Oral Squamous Cell Carcinoma. J Oral Maxillofac Surg 2019;77:536-45. [Crossref] [PubMed]
6. Pithakumar A, Basha S, Pai AR, et al. From pathogenesis to precision medicine: Targeting immune imbalance in multiple sclerosis. Ageing Res Rev 2026;113:102921. [Crossref] [PubMed]
7. Aung TN, Liu M, Su D, et al. Pathologist-Read vs AI-Driven Assessment of Tumor-Infiltrating Lymphocytes in Melanoma. JAMA Netw Open 2025;8:e2518906. [Crossref] [PubMed]
8. Assudani PJ, Bhurgy AS, Kollem S, et al. Artificial intelligence and machine learning in infectious disease diagnostics: a comprehensive review of applications, challenges, and future directions. Microchem J 2025;218:115802.
9. Zheng J, Xu X, Chen X, et al. Construction and validation of immune prognosis model for lung adenocarcinoma based on machine learning. Front Oncol 2025;15:1630663. [Crossref] [PubMed]
10. Bilal M, Nimir M, Snead D, et al. Role of AI and digital pathology for colorectal immuno-oncology. Br J Cancer 2023;128:3-11. [Crossref] [PubMed]
11. Zheng S, Wang W, Shen L, et al. Tumor battlefield within inflamed, excluded or desert immune phenotypes: the mechanisms and strategies. Exp Hematol Oncol 2024;13:80. [Crossref] [PubMed]
12. Bruni S, Mercogliano MF, Mauro FL, et al. Cancer immune exclusion: breaking the barricade for a successful immunotherapy. Front Oncol 2023;13:1135456. [Crossref] [PubMed]
13. Chakraborty S, Basak U, Mukherjee S, et al. Cancer Stem Cells Decide the Fate of Cancer Immunotherapy by Remodeling Tumor Microenvironment. Cancer Control 2025;32:10732748251381441. [Crossref] [PubMed]
14. Chong PL, Vaigeshwari V, Mohammed Reyasudin BK, et al. Integrating artificial intelligence in healthcare: applications, challenges, and future directions. Future Sci OA 2025;11:2527505. [Crossref] [PubMed]
15. Askin S, Burkhalter D, Calado G, et al. Artificial Intelligence Applied to clinical trials: opportunities and challenges. Health Technol (Berl) 2023;13:203-13. [Crossref] [PubMed]
16. Niu R, Wang C, Xie Y, et al. Prognostic significance of CD8 + tumor-infiltrating lymphocytes in operable breast cancer: a meta-analysis. BMC Cancer 2025;25:601. [Crossref] [PubMed]
17. Khoury T, Nagrale V, Opyrchal M, et al. Prognostic Significance of Stromal Versus Intratumoral Infiltrating Lymphocytes in Different Subtypes of Breast Cancer Treated With Cytotoxic Neoadjuvant Chemotherapy. Appl Immunohistochem Mol Morphol 2018;26:523-32. [Crossref] [PubMed]
18. Zou Y, Li D, Yu X, et al. Correlation of Neuroendocrine Differentiation with a Distinctively Suppressive Immune Microenvironment in Gastric Cancer. Neuroendocrinology 2024;114:192-206. [Crossref] [PubMed]