Assessing postoperative pancreatic fistula risk: from subjective assessment to digital pathology-assisted quantitative precision

Postoperative pancreatic fistula (POPF) remains one of the most feared complications following pancreatoduodenectomy, contributing substantially to postoperative morbidity, prolonged hospitalization, increased healthcare costs, and, in severe cases, mortality (1). Despite decades of technical refinement and perioperative optimization, the incidence of clinically relevant POPF has remained relatively stable. A major challenge in mitigating this complication lies in the limited precision of current risk assessment tools, many of which rely heavily on subjective intraoperative evaluation, particularly the surgeon’s assessment of pancreatic texture. Although scoring systems such as the Fistula Risk Score and its modified versions have improved perioperative risk stratification, they remain partially dependent on experiential judgment and indirect surrogates of tissue quality (2-4). Therefore, there is an unmet need for reproducible and objective predictors that can enhance individualized POPF risk estimation, beyond traditional histomorphological and tactile assessment (5-7).

In this context, the study by Màlyi et al. is the first multicenter study to correlate artificial intelligence (AI)-assisted quantified pancreatic tissue composition with POPF, marking a transition from subjective palpation to objective digital pathology (8).

The primary objective of this study was to use AI-assisted analysis to characterize pancreatic tissue composition from histological data and to correlate these findings with clinicopathological parameters from the RECOPANC (Reconstruction After Pancreatoduodenectomy) multicenter randomized controlled trial, which compared pancreatogastrostomy with pancreatojejunostomy (9). From the original 320 patients enrolled in the RECOPANC study, 134 cases met technical and quality criteria for digital analysis. Hematoxylin and eosin (H&E) stained slides from the pancreatic neck resection margin were digitized. The researchers utilized the QuPath software (Edinburgh, Scotland, UK) platform to train a machine learning algorithm to differentiate tissue compartments into three distinct categories: acinar tissue, fibrotic tissue, and adipose tissue (10). The algorithm classified the entire tissue area using training fields annotated by three pathologists, enabling the calculation of the relative and absolute areas of each tissue type. Through supervised machine learning, the investigators generated quantitative metrics of absolute and relative tissue content, which were subsequently integrated with detailed clinical and operative data.

The study presented several statistically significant and clinically relevant results regarding the micro-architecture of the pancreas and its relationship to anastomotic failure. The principal finding of the study was that relative fibrotic tissue content emerged as a predictor of clinically relevant POPF (8). The final multivariable model identified male gender [odds ratio (OR) 3.42, 95% confidence interval (CI): 1.096–10.674] and the percentage of fibrotic content (OR 0.975; 95% CI: 0.953–0.997) as independent risk factors, achieving an area under the curve (AUC) of 0.73 with 80% sensitivity and 62% specificity. While the individual AUC for fibrotic content was 0.65, the combined model demonstrated improved predictive capability. Specifically, a higher proportion of fibrosis at the resection margin was associated with a reduced risk of fistula formation, whereas higher acinar content was associated with an increased risk (univariable analyses). Importantly, digitally quantified fibrosis outperformed subjective surgical palpation in predicting POPF, highlighting the limitations of traditional texture assessment and corroborating previous findings from the RECOPANC consortium regarding the variability of clinical judgment (8,11).

The significance of this work lies in establishing the first robust multicenter evidence for histologically quantified fibrosis. Unlike previous single-center studies relying on semiquantitative visual assessment, this approach leverages the RECOPANC infrastructure to validate a standardized, reproducible framework across diverse clinical settings. Furthermore, the integration of advanced statistical methods, specifically elastic net regression and cross-validation, minimizes overfitting and lends methodological rigor to the predictive modeling. This shift from qualitative interpretation to quantitative precision marks a critical advance in pancreatic surgery risk stratification. Rather than relying on simple univariate associations, the authors used modern feature selection approaches to focus on the most relevant predictors while reducing the risk of overfitting (8). This analytic framework aligns with best practices in translational data science and strengthens the credibility of the proposed model.

More importantly, this work challenges the long-standing reliance on subjective intraoperative assessment by demonstrating that pancreatic tissue quality can be quantified in a reproducible and meaningful manner. For decades, surgeons have depended on tactile sensation and visual inspection to estimate gland fragility, despite well-recognized interobserver variability (6,11). The present study suggests that digital histopathology may offer a pathway toward standardizing this critical component of risk assessment. If validated prospectively, such an approach could serve as a foundation for integrating objective tissue metrics into existing fistula risk models, thereby refining perioperative decision-making and moving the field closer to true precision surgery (2,3).

While the study establishes a proof-of-concept for digital pathology in this domain, several limitations merit discussion, many of which are acknowledged by the authors. A substantial proportion of the original RECOPANC cohort was excluded because of missing specimens or technical artifacts, resulting in the analysis of only 134 of 320 enrolled patients (8). Although this reflects practical challenges inherent to digital pathology, particularly when working with frozen sections, it raises concerns regarding potential selection bias and generalizability (12). Patients with suboptimal specimen quality may differ systematically from those included in the final analysis, potentially influencing observed associations.

Moreover, the discriminatory performance of the predictive model, while promising, remains moderate. An AUC of approximately 0.73 indicates meaningful predictive capacity but falls short of the accuracy required for standalone clinical decision-making (8). This suggests that digital histopathology should currently be viewed as a complementary tool rather than a definitive predictor. Integration with established clinical, radiological, and biochemical parameters will likely be necessary to achieve clinically transformative performance.

Technical heterogeneity also represents an important challenge. Variations in staining protocols, section thickness, freezing artifacts, and scanning parameters across centers can influence image quality and algorithmic performance. Although the authors undertook extensive quality control and manual validation, standardization of pre-analytical and analytical workflows will be essential for widespread implementation, as emphasized in international guidelines on computational pathology (13). Future studies should prioritize harmonized protocols and external validation across independent datasets.

From a practical perspective, the feasibility of intraoperative application remains uncertain. While the authors appropriately emphasize the potential role of frozen sections in real-time risk assessment, current digital pathology pipelines may not yet be sufficiently rapid or automated for routine intraoperative use in many institutions (14). Advances in slide scanning speed, cloud-based computation, and real-time machine learning inference will be critical to translating this approach into everyday surgical workflows.

The study’s finding that fatty tissue content was not a significant predictor differs from findings by Gaujoux et al. and Mathur et al., who identified fatty infiltration as a predictor (15,16). The authors suggest this discrepancy might be due to methodology; previous studies largely relied on “analogue” visual examination rather than digital quantification. However, another possibility lies in the biological interaction between fat and fibrosis. Sugimoto et al. previously suggested that while fat infiltration is a risk, it interacts with lobularity and fibrosis (17). By isolating these variables digitally, Màlyi et al. may have revealed that the lack of fibrosis (structural integrity) is the primary driver of leakage, rather than the presence of fat per se. The structural weakness of a gland with low fibrosis likely contributes more to suture pull-through than the presence of adipocytes alone.

When situated within broader literature, the present study represents the most comprehensive digital pathology-based investigation of POPF risk to date. Earlier work by Belyaev and colleagues established the relevance of histomorphological features using conventional methods (5), while studies by Sugimoto et al. explored pancreatic stiffness, lobular structure, and fatty infiltration using both mechanical and histological approaches (17,18). Fujita et al. applied acoustic radiation force impulse imaging combined with automated tissue quantification (19), and Partelli et al. demonstrated the prognostic role of acinar content at the resection margin (20). However, these investigations were predominantly monocentric and lacked robust multivariable modeling. The RECOPANC analysis distinguishes itself through its multicenter design, quantitative methodology, and integration with validated clinical trial data, thereby setting a new benchmark for research in this domain.

Looking forward, prospective studies incorporating real-time digital pathology are needed to validate these findings in different cohorts and evaluate their impact on clinical decision-making. Multimodal models combining histological features with radiomic, proteomic, and clinical data may further enhance predictive accuracy (21-23). Additionally, embedding automated tissue analysis within electronic health record systems and surgical planning platforms could facilitate seamless risk stratification and personalized perioperative management.

The ultimate utility of this technology lies in its application during surgery. The authors note that the ideal application is intraoperative. Future research must focus on optimizing AI algorithms specifically for frozen section artifacts. One could consider a workflow where a frozen section is scanned in the operating room, and within minutes, an “AI-Score” is returned to the surgeon. This score could dictate operative strategy. For high-risk patients (low fibrosis, high acinar content), surgeons might opt for: (I) reconstruction technique: Choosing pancreatogastrostomy over pancreatojejunostomy, or perhaps carrying out total pancreatectomy in extremely high-risk scenarios; (II) drainage: implementing specific drain management protocols or omitting drains in low-risk patients; or (III) somatostatin analogues: prophylactic administration for those identified as high-risk by the digital score (24).

The RECOPANC digital pathology study is a commendable “first of many developmental steps”. It successfully challenges the subjectivity of the “surgeon’s hand” by proving that the objective percentage of fibrotic tissue is a superior predictor of POPF than tactile sensation. While the “fatty pancreas” hypothesis was not supported by this specific data set, the study illuminates the critical protective role of fibrosis. The identified limitations regarding frozen section quality and sample attrition are not failures but rather signposts directing where technical development is needed, which includes robust algorithms capable of analyzing “noisy” intraoperative slides. As we move toward an era of precision surgery, the integration of AI-assisted histology into the operating theater appears inevitable. This study provides the foundational validation that tissue composition matters and can be quantified. The next challenge is to turn this quantification into a real-time tool that guides the surgeon’s hand, ensuring that decisions regarding anastomosis and drainage are supplemented by objective quantitative data.

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-2026-1-0132/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0132/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. Bassi C, Dervenis C, Butturini G, et al. Postoperative pancreatic fistula: an international study group (ISGPF) definition. Surgery 2005;138:8-13. [Crossref] [PubMed]
2. Callery MP, Pratt WB, Kent TS, et al. A prospectively validated clinical risk score accurately predicts pancreatic fistula after pancreatoduodenectomy. J Am Coll Surg 2013;216:1-14. [Crossref] [PubMed]
3. Mungroop TH, van Rijssen LB, van Klaveren D, et al. Alternative Fistula Risk Score for Pancreatoduodenectomy (a-FRS): Design and International External Validation. Ann Surg 2019;269:937-43. [Crossref] [PubMed]
4. Verma A, Balian J, Hadaya J, et al. Machine Learning-based Prediction of Postoperative Pancreatic Fistula Following Pancreaticoduodenectomy. Ann Surg 2024;280:325-31. [Crossref] [PubMed]
5. Belyaev O, Munding J, Herzog T, et al. Histomorphological features of the pancreatic remnant as independent risk factors for postoperative pancreatic fistula: a matched-pairs analysis. Pancreatology 2011;11:516-24. [Crossref] [PubMed]
6. Hong TH, Choi JI, Park MY, et al. Pancreatic hardness: Correlation of surgeon's palpation, durometer measurement and preoperative magnetic resonance imaging features. World J Gastroenterol 2017;23:2044-51. [Crossref] [PubMed]
7. Ray S, Dhali A, Khamrui S, et al. External validation of ISGPS two-factor, four-tier classification for the prediction of pancreatic fistula after pancreaticoduodenectomy. Am J Surg 2026;251:116714. [Crossref] [PubMed]
8. Màlyi A, Bronsert P, Schilling O, et al. Postoperative pancreatic fistula risk assessment using digital pathology based analyses at the parenchymal resection margin of the pancreas - Results from the randomized multicenter RECOPANC trial. HPB (Oxford) 2025;27:393-401. [Crossref] [PubMed]
9. Keck T, Wellner UF, Bahra M, et al. Pancreatogastrostomy Versus Pancreatojejunostomy for RECOnstruction After PANCreatoduodenectomy (RECOPANC, DRKS 00000767): Perioperative and Long-term Results of a Multicenter Randomized Controlled Trial. Ann Surg 2016;263:440-9. [Crossref] [PubMed]
10. Bankhead P, Loughrey MB, Fernández JA, et al. QuPath: Open source software for digital pathology image analysis. Sci Rep 2017;7:16878. [Crossref] [PubMed]
11. Timme S, Kayser G, Werner M, et al. Surgeon vs Pathologist for Prediction of Pancreatic Fistula: Results from the Randomized Multicenter RECOPANC Study. J Am Coll Surg 2021;232:935-945.e2. [Crossref] [PubMed]
12. Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology 2004;15:615-25. [Crossref] [PubMed]
13. Abels E, Pantanowitz L, Aeffner F, et al. Computational pathology definitions, best practices, and recommendations for regulatory guidance: a white paper from the Digital Pathology Association. J Pathol 2019;249:286-94. [Crossref] [PubMed]
14. Hanna MG, Reuter VE, Ardon O, et al. Validation of a digital pathology system including remote review during the COVID-19 pandemic. Mod Pathol 2020;33:2115-27. [Crossref] [PubMed]
15. Mathur A, Pitt HA, Marine M, et al. Fatty pancreas: a factor in postoperative pancreatic fistula. Ann Surg 2007;246:1058-64. [Crossref] [PubMed]
16. Gaujoux S, Cortes A, Couvelard A, et al. Fatty pancreas and increased body mass index are risk factors of pancreatic fistula after pancreaticoduodenectomy. Surgery 2010;148:15-23. [Crossref] [PubMed]
17. Sugimoto M, Takahashi S, Kojima M, et al. In Patients with a Soft Pancreas, a Thick Parenchyma, a Small Duct, and Fatty Infiltration Are Significant Risks for Pancreatic Fistula After Pancreaticoduodenectomy. J Gastrointest Surg 2017;21:846-54. [Crossref] [PubMed]
18. Sugimoto M, Takahashi S, Kojima M, et al. What is the nature of pancreatic consistency? Assessment of the elastic modulus of the pancreas and comparison with tactile sensation, histology, and occurrence of postoperative pancreatic fistula after pancreaticoduodenectomy. Surgery 2014;156:1204-11. [Crossref] [PubMed]
19. Fujita Y, Kitago M, Abe T, et al. Evaluation of Pancreatic Fibrosis With Acoustic Radiation Force Impulse Imaging and Automated Quantification of Pancreatic Tissue Components. Pancreas 2018;47:1277-82. [Crossref] [PubMed]
20. Partelli S, Andreasi V, Schiavo Lena M, et al. The role of acinar content at pancreatic resection margin in the development of postoperative pancreatic fistula and acute pancreatitis after pancreaticoduodenectomy. Surgery 2021;170:1215-22. [Crossref] [PubMed]
21. Chen BB. Artificial intelligence in pancreatic disease. Artificial Intelligence in Medical Imaging 2020;1:19-30. [Crossref]
22. Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
23. Lee W, Park HJ, Lee HJ, et al. Deep learning-based prediction of post-pancreaticoduodenectomy pancreatic fistula. Sci Rep 2024;14:5089. [Crossref] [PubMed]
24. McMillan MT, Malleo G, Bassi C, et al. Multicenter, Prospective Trial of Selective Drain Management for Pancreatoduodenectomy Using Risk Stratification. Ann Surg 2017;265:1209-18. [Crossref] [PubMed]