Artificial intelligence and radiologists in pancreatic cancer detection using standard of care CT scans (PANORAMA): an international, paired, non-inferiority, confirmatory, observational study Alves N, Schuurmans M, Rutkowski D, et al. Lancet Oncol. 2026;27(1):116–124. https://doi.org/10.1016/S1470-2045(25)00567-4. What Did the Study Show?
This study asked a simple but important question: can an artificial intelligence (AI) system detect pancreatic cancer on CT scans as well as or better than a radiologist? To answer it, the authors designed PANORAMA: a large-scale international study involving 3,440 patients across nine centres in the Netherlands, Sweden, Norway, and the USA, and 68 radiologists from 40 centres in 12 countries. All scans were standard portal-venous phase contrast-enhanced CTs. Every diagnosis was confirmed either by tissue biopsy or at least three years of clinical follow-up. The AI system was an ensemble of the top three algorithms from an open international challenge attracting 432 developers across 46 countries. It achieved an AUROC of 0.92 on the sequestered test cohort of 1,130 patients. In the matched reader sub-study of 391 patients, it was statistically superior to the pooled radiologist performance (AI AUROC 0.92 vs radiologists 0.88; p=0.001). Translated into clinical terms: at the same sensitivity level, the AI produced 38% fewer false positives than radiologists. What makes this study stand out?
Most previous AI studies in this field used smaller, single-centre, proprietary datasets. None had made their algorithms openly available. PANORAMA pre-registered its protocol and statistical plan, and has released the trained algorithm, full training dataset, and analysis code publicly. This level of transparency is commendable in AI research and represents a step towards open science. Are there any caveats or limitations?
Data mainly originated from European tertiary centres; the AI was tested as a standalone tool rather than alongside a radiologist; the reader study was conducted in a controlled online environment without access to clinical notes or prior imaging; the enriched test cohort PDAC prevalence (37%) substantially exceeds what would be encountered in routine practice. What Does This Mean for Radiologists?
These results should be read not as a challenge to radiologists, but as a signal of where AI can add the most value. PDAC detection on CT is notoriously difficult: secondary signs are subtle, small early-stage lesions are frequently missed, and inter-reader variability is high. In practice, AI could be integrated into the reporting workflow as a pre-reader flagging suspicious pancreatic morphology before the radiologist reports, or as a concurrent second reader drawing attention to subtle pancreatic abnormalities that might otherwise be underscored on a busy list. In a real-world assisted reading scenario, the radiologist remains the interpreter, integrating AI output with clinical history, prior imaging, and laboratory data in a way a standalone system cannot. This human-AI synergy is likely where the greatest diagnostic gains lie, though it remains to be tested prospectively. For ESGAR members working in abdominal imaging, the open-source benchmark also invites independent validation across European centres, something the authors have committed to maintaining for at least another five years. What Does This Mean for Patients?
Pancreatic cancer is the deadliest of the major cancers, with a median overall survival of just four months across all disease stages and over 467,000 annual deaths worldwide. Patients with early-stage resectable disease, however, have a median survival of 32 months, making early detection the single most impactful lever available. Beyond improved sensitivity, the reduction in false positives is equally meaningful for patients: unnecessary referrals carry real costs in anxiety, radiation exposure, and resource use, particularly if AI is eventually deployed at scale for opportunistic screening on routine imaging. Can This Be Easily Implemented, and What Are the Next Steps?
Implementation is feasible but not yet straightforward. Performance in community hospitals, where volumes, expertise, and CT protocols differ, remains untested. Practical deployment will also require seamless integration within existing PACS and radiology IT infrastructure, a non-trivial technical and organisational challenge that has yet to be addressed. Future work must also extend beyond PDAC vs normal to reflect the broader differential diagnosis encountered in clinical practice. The authors highlight opportunistic screening as a particularly compelling next application, proposing the detection of early asymptomatic PDAC on routine CT without added cost or radiation, although this requires dedicated prospective evaluation. PANORAMA represents a meaningful methodological step forward; further evidence will be needed to confirm its role in clinical practice. | References Alves N, Schuurmans M, Rutkowski D, et al. PANORAMA. Lancet Oncol. 2026;27(1):116–124. Conroy T, Pfeiffer P, Vilgrain V, et al. Pancreatic cancer: ESMO Clinical Practice Guideline. Ann Oncol. 2023;34:987–1002. 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. Toft J, Hadden WJ, Laurence JM, et al. Imaging modalities in the diagnosis of pancreatic adenocarcinoma. Eur J Radiol. 2017;92:17–23. Saha A, Bosma JS, Twilt JJ, et al. AI and radiologists in prostate cancer detection on MRI (PI-CAI). Lancet Oncol. 2024;25:879–87. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022. CA Cancer J Clin. 2024;74:229–63. Stoop TF, Javed AA, Oba A, et al. Pancreatic cancer. Lancet. 2025;405:1182–202. Luyer MDP. AI in pancreatic cancer detection: from premise to practice. Lancet Oncol. 2026;27(1):8–9.
|
