Pancreatic ductal adenocarcinoma (PDAC) remains one of the deadliest cancers worldwide, with rising incidence and survival that has barely improved over the past decade. Early detection is notoriously challenging: small tumors are subtle, easily missed, and often diagnosed only when curative options are no longer possible. Against this backdrop, the PANORAMA study was developed to address one of the most urgent questions in gastrointestinal oncology—whether artificial intelligence (AI) can reliably support or even outperform radiologists in detecting PDAC on routine contrast-enhanced CT, the global first-line imaging modality for suspected pancreatic cancer.
Although CT is indispensable for diagnosis and staging, its limitations are well known. Early PDAC can mimic normal pancreatic tissue, inter-reader variability is substantial, and secondary signs may be overlooked even by experienced radiologists. Meanwhile, AI has demonstrated expert-level performance in several other cancer-imaging domains, yet robust evidence supporting its role in PDAC has been scarce. The PANORAMA study was designed to fill this critical gap through the largest paired, international evaluation of AI versus radiologists to date.
PANORAMA Study Methods and Endpoints
PANORAMA included 3440 adults across five European tertiary centers, complemented by publicly available datasets from the USA. All CT examinations were performed in the portal-venous phase, and all PDAC cases in the tuning and sequestered testing cohorts were histologically confirmed. Negative control cases required histology or at least 3 years of clinical follow-up without PDAC.
The study consisted of two coordinated arms: an AI grand challenge and an international reader study.
- The AI system was developed from a public dataset of 2224 patients, with independent evaluation on a sequestered cohort of 1130 patients (406 PDAC). The final AI system was an ensemble of the three top-performing algorithms (Siemens Healthineers, Cedars-Sinai, Guerbet Research).
- The reader study enrolled 68 radiologists from 12 countries, each interpreting a randomly selected subset of 391 cases (144 PDAC).
Readers gave a PDAC likelihood score (0–100) and a 0–5 Likert score; no clinical history or prior imaging was provided.
The primary endpoint was the difference in AUROC between the AI system and the pooled radiologists. The study used a hierarchical testing strategy: first assessing non-inferiority (margin 0.05), and only if met, proceeding to test for superiority. Secondary analyses evaluated sensitivity, specificity, and false-positive/false-negative patterns at clinically relevant thresholds derived from the Likert scoring system.
You can Read about How Artificial Intelligence Is Transforming Cancer Care in 2025: Diagnosis, Treatment, Clinical Trials, and Screening on OncoDaily.
Results of PANORAMA study
The results of the PANORAMA study were published in The Lancet Oncology on November 20, 2025. In the sequestered testing cohort of 1130 patients, the AI system achieved an AUROC of 0.92 (95% CI 0.90–0.93), with a sensitivity of 85.7% and a specificity of 83.5%. In the paired reader study, the pooled radiologist AUROC was 0.88 (95% CI 0.85–0.91), while AI again achieved 0.92 (95% CI 0.89–0.94). AI met the prespecified non-inferiority margin of 0.05 (p < 0.0001) and subsequently showed statistical superiority (p = 0.001).
At a highly sensitive operating point comparable to a Likert score of 1, radiologists reached 96.1% sensitivity but only 44.2% specificity. When calibrated to the same sensitivity, AI produced 38% fewer false positives (85 versus 138). At a threshold approximating routine clinical CT interpretation (similar to Likert ≥2), AI missed 13 cancers while radiologists missed 21 at the same specificity of 73.8%. When matched for sensitivity (85.6%), AI generated 48 false positives compared with 65 for radiologists, representing a 26% reduction.
A post-hoc review showed that none were normal scans; all five had been considered suspicious in the original clinical radiology reports. Each represented significant pancreatic or peri-pancreatic disease, including pseudocyst, cystadenoma, cholangiocarcinoma invading the pancreas, metastatic colon cancer, and duodenal carcinoma—supporting the clinical appropriateness of these alerts.
Conclusion
The PANORAMA study shows that AI, when trained on large and diverse datasets, can exceed average radiologist performance in detecting pancreatic cancer on routine CT. It reduced both missed cancers and avoidable work-ups across key clinical thresholds and provided a transparent, open benchmark that sets a new standard for AI research in PDAC.
Future prospective trials will determine how AI can be integrated into daily radiology workflows, accelerate early diagnosis, and ultimately improve outcomes in one of oncology’s most challenging diseases.
You can also read about AI in Breast Cancer Detection: What Patients Should Know on OncoDaily.