Artificial intelligence detection of pancreatic cancer: lessons from the PANORAMA study

Introduction

Pancreatic ductal adenocarcinoma (PDAC) remains uniquely deadly, with survival rates that lag far behind other solid malignancies. This disparity is driven largely by late detection: the majority of PDAC patients present with late-stage disease which is no longer eligible for surgical resection (1). Because earlier detection of PDAC translates into longer survival, there is a critical need to identify smaller tumors before they become locally advanced or metastasize. Unfortunately, PDAC is most often detected after the onset of nonspecific or obstructive symptoms, which typically occur late in the disease course and prompt diagnostic imaging. Although contrast-enhanced computed tomography (CT) is the primary modality for PDAC detection, imaging findings are often subtle and subject to substantial inter-reader variability. Artificial intelligence (AI) has the potential to identify subtle pancreatic abnormalities on CT, helping radiologists avoid missed detections that might otherwise only be recognized retrospectively. The PANORAMA study used AI to detect pancreatic cancer from contrast-enhanced CT images in a large international multicenter cohort (2). Compellingly, PANORAMA found that AI was not only non-inferior but outperformed a panel of 68 radiologists in PDAC detection. In this commentary we discuss the strengths and limitations of the PANORAMA study and the implications for pancreatic cancer screening using CT.

Role of CT in PDAC detection

Contrast-enhanced CT serves as the first-line imaging modality for the detection of PDAC, owing to its diagnostic capability, widespread availability, and comparatively low cost. Current National Comprehensive Cancer Network (NCCN) guidelines recommend a dedicated dual-phase pancreatic protocol, with image acquisition during both the late arterial and portal venous phases to optimize tumor conspicuity and vascular assessment (3). Magnetic resonance imaging (MRI), while valuable, is generally reserved for problem-solving scenarios, such as when clinical suspicion for PDAC persists despite a normal CT examination.

On CT, PDAC typically manifests as an ill-defined, hypoattenuating mass, classically located in the pancreatic head and often accompanied by a surrounding desmoplastic reaction (4). Due to its hypovascularity, PDAC demonstrates minimal contrast enhancement and remains hypoattenuating relative to the adjacent pancreatic parenchyma on both arterial and venous phase imaging, though it is typically more conspicuous during the arterial phase. Secondary imaging features include upstream dilatation of the main pancreatic duct and, in tumors arising in the pancreatic head, the “double duct sign”, reflecting simultaneous dilatation of the pancreatic and common bile ducts due to malignant obstruction.

Clinical indications for contrast-enhanced CT in the evaluation of suspected PDAC include classic warning symptoms such as painless jaundice, unexplained abdominal or back pain, unintentional weight loss, fatigue, and anorexia (3). CT is also warranted in high-risk clinical scenarios, particularly new-onset diabetes mellitus or recurrent episodes of idiopathic pancreatitis, both of which may represent early manifestations of an underlying pancreatic malignancy (3). In addition, incidental findings on prior imaging—such as focal pancreatic abnormalities, main pancreatic duct dilatation, or abrupt ductal cutoff—should prompt dedicated pancreatic protocol CT to further characterize these findings and evaluate for occult neoplasia.

Despite the central role of CT in PDAC detection, population-wide screening is not recommended. The U.S. Preventive Services Task Force advises against screening asymptomatic adults without known high-risk features, given the low prevalence of disease and potential harms of false-positive findings (5). Surveillance is instead reserved for individuals at substantially elevated risk, including those with a strong familial predisposition (typically defined as two or more first-degree relatives with pancreatic cancer) and carriers of pathogenic variants in established pancreatic cancer susceptibility genes (6). In these high-risk populations, the current NCCN guidelines recommend screening with magnetic resonance cholangiopancreatography (MRCP) and/or endoscopic ultrasound (EUS), both of which offer superior sensitivity for detecting small or early pancreatic lesions compared with CT.

Once PDAC is identified on CT, a structured workup is initiated to ensure accurate staging and guide management. When the initial diagnosis is made on a standard CT examination, a dedicated pancreatic protocol CT should be obtained to better characterize tumor extent and evaluate vascular involvement (3,7). Comprehensive staging with CT imaging of the chest and pelvis is also recommended to assess for metastatic disease. Serum carbohydrate antigen 19-9 (CA19-9) and liver function tests, while not diagnostic in isolation, provide important prognostic context and a baseline against which to monitor treatment response (8). All newly diagnosed cases warrant multidisciplinary review to assess resectability and guide treatment planning. Finally, patients should be offered germline genetic testing to identify inherited susceptibility mutations, and, when feasible, tumor molecular profiling should be performed to help inform potential targeted therapeutic strategies (8).

PANORAMA study strengths

Despite the growing volume of literature exploring the role of AI in diagnostic radiology, relatively few studies meet the methodological standards needed to meaningfully inform real-world practice. The PANORAMA study distinguishes itself in this regard, incorporating several design strategies that strengthen the validity and credibility of its findings.

Chief among these is its pre-registered, confirmatory study design, which contrasts with the predominantly exploratory nature of many prior studies exploring AI applications in radiology. Rather than retrospectively reporting AI performance without a prespecified hypothesis, the investigators defined a clear primary question from the very beginning: whether the AI system would prove non-inferior to radiologists in the detection of PDAC on CT. Moreover, the pre-registration of primary and secondary endpoints aligns with best practices for diagnostic accuracy research (9) and reduces the risk for selective reporting or post-hoc manipulation of endpoint thresholds.

The decision to develop the AI model through a publicly hosted challenge framework was equally consequential. By providing open access to a shared, multicenter CT dataset, the investigators enabled teams of developers worldwide to independently build and refine PDAC detection algorithms, promoting methodological diversity and drawing on the expertise of international experts. This commitment to openness extended beyond the challenge itself: the finalized AI algorithm, source code, full training dataset, and benchmarking platform have all been made publicly available, facilitating reproducibility and promoting scientific collaboration.

Once finalized, the AI system was tested on a fully sequestered dataset that was completely independent of the training and tuning data, thus ensuring that each examination was encountered for the first time at evaluation. This design more closely mirrors real-world deployment and mitigates the risk of inflated performance estimates, thereby strengthening the generalizability of the reported results. Such external validation is critical, as AI models frequently exhibit performance degradation when applied to patient populations outside the training data (10).

The diversity of the underlying imaging data further supports the generalizability of the PANORAMA study’s findings. Both the training and evaluation datasets were drawn from multiple tertiary care centers across several countries, addressing a major limitation of prior AI studies in pancreatic cancer detection, which have largely relied on single-center cohorts. The inclusion of CT scans acquired from a range of vendors including Canon, GE, Siemens, Philips, Toshiba, and Picker, adds another layer of robustness and supports the AI model’s ability to perform outside tightly controlled conditions.

The scale of the study represents another notable strength. Whereas most earlier investigations of AI-based pancreatic cancer have relied on cohorts of fewer than 1000 patients, PANORAMA draws on a substantially larger pool, providing a more robust foundation for model development and evaluation. Equally important is the scope of the reader study, which included 68 radiologists from 12 countries with a median of nine years of experience. This level of radiologist participation far exceeds that of most reported studies and offers a more representative and credible benchmark against which to assess AI performance.

Finally, the PANORAMA study is strengthened by its use of a high-quality reference standard. While many prior AI studies have relied primarily on radiology reports alone for model training, all PDAC-positive cases in PANORAMA were confirmed histologically. Negative control cases were also strictly defined, requiring either histologic confirmation or a minimum of three years of clinical follow-up without subsequent PDAC diagnosis. Together, these measures minimize the risk of misclassification and lend substantial credibility to the reported performance of both the AI system and the radiologists (9).

PANORAMA study limitations

Despite its methodological rigor, several important limitations warrant consideration when interpreting PANORAMA’s findings and planning next steps toward clinical translation. One of the most important considerations is the enrichment of PDAC prevalence in the reader study subset. In this cohort, 37% of patients had histologically confirmed PDAC, which is considerably higher than the prevalence encountered in high-risk surveillance populations, ranging from approximately 1.56% in familial high-risk cohorts to as high as 7.3% among CDKN2A mutation carriers (6,11-13). Because positive predictive value is strongly influenced by disease prevalence, the AI system’s reported specificity of 84.6% would be expected to perform differently at real-world prevalence. This model trained on enriched training data will likely display lower specificity in the clinic, with a higher number of false positives. The downstream implications of unnecessary EUS-guided biopsies, patient anxiety, and healthcare costs at such false positive rates would pose significant challenges for clinical implementation (14). As the field targets earlier detection in asymptomatic patients, explicit modeling of AI performance at clinically realistic prevalence levels would strengthen the case for deployment and represents an important area for future analysis.

Another important consideration is the controlled conditions under which radiologists were evaluated. Readers interpreted examinations without access to clinical history, prior imaging, laboratory data, or peer consultation. In clinical practice, radiology is inherently collaborative. Challenging cases are routinely discussed with colleagues at the workstation and through multidisciplinary tumor board conferences. Second-opinion radiology consultations have been shown to alter clinical management in a substantial proportion of cases and improve diagnostic accuracy (15). AI tools trained using both imaging and non-imaging data are expected to improve cancer prediction (16). A complementary study design testing radiologist performance under conditions that more closely mirror clinical practice, including access to clinical context and peer consultation, would offer additional insight into the practical value of AI assistance.

A notable gap in the current analysis is the absence of any evaluation of the AI system’s detection maps. The Methods section states that algorithms were required to produce both a likelihood score and “a detection map localising suspicious lesions”. Yet neither the main manuscript nor the appendix presents any visualization or quantitative evaluation of these localization outputs. This represents an important opportunity for future work which will provide context and validation to this study. For clinical integration, understanding where the AI identifies suspicious features is as important as the probability score itself. Saliency-based explainability methods have been identified as one of the most important factors in building radiologist trust and fostering clinical adoption of AI (17,18). Without insight into whether the AI localizes tumors accurately or instead relies on indirect features such as upstream duct dilation, radiologists would have difficulty incorporating AI outputs into their diagnostic reasoning. This consideration takes on added relevance given that the European Union AI Act explicitly classifies medical AI as high-risk and calls for rigorous evaluation of explainability (19).

Additional considerations include the preponderance of large tumors in the study cohort (median 30.7 mm, with only 6% classified as T1 stage), which limits generalizability to the very population where AI could have the greatest clinical impact, namely small, early-stage tumors most amenable to curative resection. This limitation is underscored by a prior study of PDAC, in which AI detection sensitivity decreased for tumors under 2 cm, ranging from 63% to 80% sensitivity (20-22). Readers were also primed for a PDAC detection task, which may have elevated both AI and reader performance relative to clinical practice, where PDAC may present as a subtle incidental finding on imaging obtained for unrelated indications (23).

Implications for the radiologist and the path forward

PANORAMA provides compelling evidence that AI has matured to the point of matching average radiologist performance for PDAC detection on contrast-enhanced CT. This is a meaningful milestone, but it is best understood not as a mandate for replacement but rather as an invitation to rethink how AI can augment clinical workflows in ways that genuinely improve patient outcomes.

The most promising near-term application is as a triaging and second-reader tool, particularly in community hospitals and emergency departments where subspecialty abdominal radiology expertise may not be readily available. In such settings, AI could help reduce the well-documented variability in PDAC detection while preserving the radiologist’s central role as the interpreting physician (24). Realizing this potential, however, requires several important next steps.

First, prospective validation in real-world clinical settings is essential. Reader-assistance studies in which radiologists interpret cases with and without AI are needed to determine whether AI genuinely improves diagnostic performance in practice or introduces the risk of automation bias (25). Second, explainability should be treated as a core design requirement rather than an afterthought. The PANORAMA consortium is well positioned to analyze and publish the detection map data already collected, as radiologists need spatial context to meaningfully integrate AI predictions into their clinical reasoning. Räz et al. recently proposed that medical AI explanations should be context-dependent and user-dependent to be useful in clinical workflows (26). Third, performance should be characterized at clinically relevant disease prevalence. The PANDA study’s validation of over 18,000 consecutive patients offers a more clinically realistic benchmark for evaluating AI performance in an unselected screening population (27). Fourth, equity and generalizability deserve attention. PANORAMA’s predominantly European dataset and limited demographic data leave open the possibility of performance differences across underrepresented populations, a concern supported by recent evidence of demographic bias in medical imaging AI (28,29).

Ultimately, the field must move beyond the question of whether AI can detect PDAC to the more consequential questions of how, where, and for whom it should be deployed. The regulatory pathway will require prospective trials measuring meaningful clinical endpoints, including time to diagnosis, resectability rates, and patient survival, rather than relying solely on area under the receiver operating characteristic (AUROC) curve derived from sequestered datasets. The radiologist is not being replaced, but the tools, workflows, and expectations surrounding diagnostic imaging are evolving. PANORAMA lays important groundwork, and the next phase of clinical translation will build on it.

Acknowledgments

None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Journal of Medical Artificial Intelligence. The article has undergone external peer review.

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-1-0046/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-1-0046/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. Singhi AD, Koay EJ, Chari ST, et al. Early Detection of Pancreatic Cancer: Opportunities and Challenges. Gastroenterology 2019;156:2024-40. [Crossref] [PubMed]
2. Alves N, Schuurmans M, Rutkowski D, et al. Artificial intelligence and radiologists in pancreatic cancer detection using standard of care CT scans (PANORAMA): an international, paired, non-inferiority, confirmatory, observational study. Lancet Oncol 2026;27:116-24. [Crossref] [PubMed]
3. NCCN Clinical Practice Guidelines in Oncology: Pancreatic Adenocarcinoma. Plymouth Meeting, PA: National Comprehensive Cancer Network; 2025.
4. Webb WR, Brant WE, Major NM. Fundamentals of Body CT. 3 edition. Philadelphia, PA: Elsevier/Saunders; 2026.
5. Henrikson NB, Aiello Bowles EJ, Blasi PR, et al. Screening for Pancreatic Cancer: Updated Evidence Report and Systematic Review for the US Preventive Services Task Force. JAMA 2019;322:445-54. [Crossref] [PubMed]
6. Aslanian HR, Lee JH, Canto MI. AGA Clinical Practice Update on Pancreas Cancer Screening in High-Risk Individuals: Expert Review. Gastroenterology 2020;159:358-62. [Crossref] [PubMed]
7. Kulkarni NM, Hough DM, Tolat PP, et al. Pancreatic adenocarcinoma: cross-sectional imaging techniques. Abdom Radiol (NY) 2018;43:253-63. [Crossref] [PubMed]
8. Park W, Chawla A, O'Reilly EM. Pancreatic Cancer: A Review. JAMA 2021;326:851-62. [Crossref] [PubMed]
9. Sounderajah V, Ashrafian H, Golub RM, et al. Developing a reporting guideline for artificial intelligence-centred diagnostic test accuracy studies: the STARD-AI protocol. BMJ Open 2021;11:e047709. [Crossref] [PubMed]
10. Rajpurkar P, Lungren MP. The Current and Future State of AI Interpretation of Medical Images. N Engl J Med 2023;388:1981-90. [Crossref] [PubMed]
11. Hesami Z, Olfatifar M, Sadeghi A, et al. Global Trend in Pancreatic Cancer Prevalence Rates Through 2040: An Illness-Death Modeling Study. Cancer Med 2024;13:e70318. [Crossref] [PubMed]
12. Pereira SP, Oldfield L, Ney A, et al. Early detection of pancreatic cancer. Lancet Gastroenterol Hepatol 2020;5:698-710. [Crossref] [PubMed]
13. Vasen H, Ibrahim I, Ponce CG, et al. Benefit of Surveillance for Pancreatic Cancer in High-Risk Individuals: Outcome of Long-Term Prospective Follow-Up Studies From Three European Expert Centers. J Clin Oncol 2016;34:2010-9. [Crossref] [PubMed]
14. Gonda TA, Everett JN, Wallace M, et al. Recommendations for a More Organized and Effective Approach to the Early Detection of Pancreatic Cancer From the PRECEDE (Pancreatic Cancer Early Detection) Consortium. Gastroenterology 2021;161:1751-7. [Crossref] [PubMed]
15. Abujudeh HH, Boland GW, Kaewlai R, et al. Abdominal and pelvic computed tomography (CT) interpretation: discrepancy rates among experienced radiologists. Eur Radiol 2010;20:1952-7. [Crossref] [PubMed]
16. Haue AD, Hjaltelin JX, Holm PC, et al. Artificial intelligence-aided data mining of medical records for cancer detection and screening. Lancet Oncol 2024;25:e694-703. [Crossref] [PubMed]
17. Pertuz S, Ortega D, Suarez É, et al. Saliency of breast lesions in breast cancer detection using artificial intelligence. Sci Rep 2023;13:20545. [Crossref] [PubMed]
18. Borys K, Schmitt YA, Nauta M, et al. Explainable AI in medical imaging: An overview for clinical practitioners - Beyond saliency-based XAI approaches. Eur J Radiol 2023;162:110786. [Crossref] [PubMed]
19. Union EPaCotE. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union; 2024.
20. Chen PT, Wu T, Wang P, et al. Pancreatic Cancer Detection on CT Scans with Deep Learning: A Nationwide Population-based Study. Radiology 2023;306:172-82. [Crossref] [PubMed]
21. Korfiatis P, Suman G, Patnam NG, et al. Automated Artificial Intelligence Model Trained on a Large Data Set Can Detect Pancreas Cancer on Diagnostic Computed Tomography Scans As Well As Visually Occult Preinvasive Cancer on Prediagnostic Computed Tomography Scans. Gastroenterology 2023;165:1533-1546.e4. [Crossref] [PubMed]
22. Alves N, Schuurmans M, Litjens G, et al. Fully Automatic Deep Learning Framework for Pancreatic Ductal Adenocarcinoma Detection on Computed Tomography. Cancers (Basel) 2022;14:376. [Crossref] [PubMed]
23. Gangi S, Fletcher JG, Nathan MA, et al. Time interval between abnormalities seen on CT and the clinical diagnosis of pancreatic cancer: retrospective review of CT scans obtained before diagnosis. AJR Am J Roentgenol 2004;182:897-903. [Crossref] [PubMed]
24. Al-Hawary MM, Francis IR, Chari ST, et al. Pancreatic ductal adenocarcinoma radiology reporting template: consensus statement of the society of abdominal radiology and the american pancreatic association. Gastroenterology 2014;146:291-304.e1. [Crossref] [PubMed]
25. Tschandl P, Rinner C, Apalla Z, et al. Human-computer collaboration for skin cancer recognition. Nat Med 2020;26:1229-34. [Crossref] [PubMed]
26. Räz T, Pahud De Mortanges A, Reyes M. Explainable AI in medicine: challenges of integrating XAI into the future clinical routine. Front Radiol 2025;5:1627169. [Crossref] [PubMed]
27. Cao K, Xia Y, Yao J, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med 2023;29:3033-43. [Crossref] [PubMed]
28. Yang Y, Liu Y, Liu X, et al. Demographic bias of expert-level vision-language foundation models in medical imaging. Sci Adv 2025;11:eadq0305. [Crossref] [PubMed]
29. Seyyed-Kalantari L, Zhang H, McDermott MBA, et al. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med 2021;27:2176-82. [Crossref] [PubMed]