AI Detects Occult Pancreatic Cancer 15 Months Before Radiologists on CT

The challenge of intercepting occult pancreatic malignancy

Pancreatic ductal adenocarcinoma remains one of the most lethal solid malignancies, with 2025 estimates projecting over 600,000 cancer deaths in the United States alone [1]. Long-term survival is primarily contingent on early surgical resection with macroscopic tumor clearance, yet resectability rates drop from 73.6% (95% CI, 65.9 to 80.6) in localized disease to 33.2% (95% CI, 25.8 to 41.1) in borderline or unresectable cases [2]. Standard contrast-enhanced computed tomography often misses visually occult lesions, and while serum carbohydrate antigen 19-9 (a tumor-associated glycoprotein) is a specific marker for malignancy at 0.90, its sensitivity remains inadequate for early screening at 0.38 [3, 4, 5]. Recent meta-analyses of artificial intelligence models using radiomics (the extraction of quantitative data from clinical images) demonstrate a pooled sensitivity of 0.88 (95% CI, 0.84 to 0.91) and specificity of 0.93 (95% CI, 0.87 to 0.96) for detecting these early lesions [6, 7]. To address this diagnostic gap, researchers recently evaluated an automated framework designed to identify subvisual architectural disruptions on standard imaging before they manifest as detectable masses, potentially offering a critical window for earlier surgical intervention.

Automated detection of subvisual architectural disruptions

The researchers developed and validated the Radiomics-based Early Detection MODel (REDMOD), an automated framework designed to identify subvisual radiomic signatures of pre-diagnostic pancreatic ductal adenocarcinoma on standard-of-care computed tomography scans. The system relies on radiomics, a technique that extracts high-throughput quantitative data from medical images to detect structural patterns invisible to the human eye. To build this model, the study utilized a multi-institutional training cohort of 969 participants, which included 156 pre-diagnostic cases and 813 controls. This large dataset provided the necessary breadth to capture the subtle variations in pancreatic tissue that precede clinical diagnosis. The REDMOD framework employs artificial intelligence-driven segmentation, automatically identifying and outlining the boundaries of the pancreas, coupled with a heterogeneous ensemble architecture that combines multiple machine learning algorithms to improve predictive accuracy. To address the challenge of training a model on a condition with low clinical prevalence, the researchers used the Synthetic Minority Over-sampling Technique (SMOTE), a statistical method that generates synthetic examples of the minority class to balance the dataset. This approach allowed the model to learn from a 40-feature radiomic signature without being biased toward the more common control cases. Mechanistic analyses revealed that the predictive power of the model was derived principally from multi-scale wavelet-filtered textural features, which comprised 90% of the selected 40-feature signature. Wavelet filtering is a mathematical method used to decompose images into different scales and frequencies, enabling the identification of hidden patterns at various levels of detail. These wavelet-filtered features were significantly more effective at capturing subvisual architectural disruptions than unfiltered features, achieving an area under the curve of 0.82 compared to 0.74 for unfiltered data (p=0.007). By isolating these specific textural changes, the model can detect the earliest signs of malignancy before a discrete mass becomes visible, offering clinicians a tool to flag high-risk patients for closer surveillance.

Diagnostic performance and lead time advantages

The researchers evaluated the model using an independent test set of 493 participants, which included 63 pre-diagnostic cases and 430 controls. This testing phase was specifically designed to simulate a low prevalence early detection paradigm of approximately 1:6, reflecting the clinical reality of screening for pancreatic ductal adenocarcinoma in high-risk populations. Within this cohort, the model achieved an area under the curve of 0.82, indicating a strong overall ability to distinguish between patients who would develop cancer and those who would not. To ensure clinical utility, the framework utilizes a tunable Youden Index-optimized classification threshold, a statistical method used to define the optimal cutoff point for a positive test result. This feature allows clinicians to calibrate the model's performance for different screening goals, such as prioritizing high sensitivity in a high-risk genetic cohort, without the need for retraining the underlying algorithm. In direct comparisons with clinical experts, the model demonstrated a significant advantage in identifying malignancy during the visually occult stage. At a median lead time of 475 days before clinical diagnosis, the model achieved a 73.0% sensitivity, identifying nearly twice as many cases as radiologists, who reached a sensitivity of 38.9% (p < 0.001). This diagnostic gap widened significantly when looking further back in the disease progression. At a lead time of greater than 24 months before diagnosis, the model maintained a 68.0% sensitivity, representing a nearly threefold increase over the 23.0% sensitivity achieved by radiologists. For practicing physicians, these data suggest that the automated system can detect subtle textural changes in pancreatic tissue more than two years before a discrete mass becomes visible on standard imaging, potentially shifting the timeline of diagnosis to a highly resectable stage.

Reliability across institutions and timeframes

For a diagnostic tool to be viable in a clinical setting, it must demonstrate consistent results over time and across different imaging environments. The researchers addressed this by performing a longitudinal test-retest analysis, which measures the consistency of the model's output when the same patient is scanned at different intervals. The model demonstrated longitudinal stability with 90% to 92% concordance, indicating that the algorithm provides highly consistent classifications across repeated imaging sessions. This level of reliability is essential for clinicians who may use these scores to monitor high-risk patients over several years, as it ensures that changes in the risk score reflect actual physiological shifts rather than technical noise or variability in image acquisition. To ensure the findings were not limited to the initial training environment, the validation process included external specificity validation across two independent cohorts. Specificity, the ability of the test to correctly identify those without the disease, was tested in a multi-institutional dataset of 539 participants and a public dataset of 80 participants. In these external groups, the model maintained an external specificity of 81.3% in the multi-institutional cohort (n=539) and 87.5% in the public dataset (n=80). These figures suggest that the tool maintains a low rate of false positives even when applied to imaging data from diverse clinical sources and different computed tomography scanner manufacturers, addressing a common concern regarding the generalizability of artificial intelligence diagnostics. The combination of high sensitivity at long lead times and stable specificity across institutions suggests a clear role for this technology in managing high-risk populations, such as patients with certain genetic predispositions or new-onset diabetes. By identifying signatures of pancreatic ductal adenocarcinoma up to 475 days before they become visible to the human eye, clinicians may be able to intercept the disease proactively, translating to a higher rate of patients eligible for curative surgical resection.

References

1. Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics, 2025. CA A Cancer Journal for Clinicians. 2025. doi:10.3322/caac.21871

2. Gillen S, Schuster T, Büschenfelde CMZ, Frieß H, Kleeff J. Preoperative/Neoadjuvant Therapy in Pancreatic Cancer: A Systematic Review and Meta-analysis of Response and Resection Percentages. PLoS Medicine. 2010. doi:10.1371/journal.pmed.1000267

3. Yan G, Chen X, Wang Y. Diagnostic performance of artificial intelligence based on contrast-enhanced computed tomography in pancreatic ductal adenocarcinoma: a systematic review and meta-analysis.. Abdominal radiology (New York). 2026. doi:10.1007/s00261-025-05089-2

4. Sultana A, Jackson R, Tim G, et al. What Is the Best Way to Identify Malignant Transformation Within Pancreatic IPMN: A Systematic Review and Meta-Analyses. Clinical and Translational Gastroenterology. 2015. doi:10.1038/ctg.2015.60

5. Yang J, Huang J, Zhang Y, et al. Contrast-enhanced ultrasound and contrast-enhanced computed tomography for differentiating mass-forming pancreatitis from pancreatic ductal adenocarcinoma: a meta-analysis.. Chinese medical journal. 2023. doi:10.1097/CM9.0000000000002300

6. Alidina Z, Hussain AAM, Banani I, Khan MM, Pawlik TM. Radiomics for early detection of pancreatic cancer: a systematic review and meta-analysis.. Journal of gastrointestinal surgery : official journal of the Society for Surgery of the Alimentary Tract. 2026. doi:10.1016/j.gassur.2026.102374

7. Harandi H, Gouravani M, Alikarami S, et al. Diagnostic Performance of Artificial Intelligence in Detecting and Distinguishing Pancreatic Ductal Adenocarcinoma via Computed Tomography: A Systematic Review and Meta-Analysis.. Journal of imaging informatics in medicine. 2026. doi:10.1007/s10278-025-01607-2