Pancreatic ductal adenocarcinoma (PDAC) is the deadliest solid cancer, primarily because it is almost always detected too late for surgery. Early detection could dramatically improve survival, but screening asymptomatic individuals has been considered impractical due to low disease prevalence and the risk of false positives.
Non-contrast CT scans are routinely performed for many clinical reasons — lung cancer screening, emergency workups, annual check-ups — creating millions of opportunities to detect pancreatic cancer early. However, detecting PDAC on non-contrast CT has long been considered nearly impossible because tumors are very difficult to see without contrast dye enhancement.
The researchers developed PANDA (Pancreatic cancer detection with Artificial intelligence), trained on 3,208 patients from a single center. PANDA works in three stages: first localizing the pancreas, then detecting subtle lesions within it, and finally classifying the type of lesion (PDAC or one of seven non-PDAC subtypes).
The model was supervised with pathology-confirmed labels and pixel-wise annotations, with lesion masks transferred from contrast CT scans where tumors are more visible. This innovative approach allowed the AI to learn what subtle non-contrast CT changes look like even though those changes are nearly imperceptible to human radiologists.
In a multicenter validation involving 6,239 patients across 10 centers, PANDA achieved an AUC of 0.984 to 0.996 for lesion detection with sensitivity of 93.3% and specificity of 98.8%. For PDAC specifically, the detection rate was 96.5%.
In a reader study comparing PANDA against 33 radiologists reviewing the same non-contrast CTs, PANDA outperformed the mean radiologist performance by 34.1% in sensitivity and 6.3% in specificity. Even compared to radiologists reading contrast-enhanced CT, PANDA on non-contrast CT remained competitive.
PANDA was deployed in a real-world clinical setting involving 20,530 consecutive patients across physical exam, emergency, outpatient, and inpatient scenarios. In one study, PANDA detected cancers missed by the standard of care, with several patients later confirmed to have PDAC by contrast MRI and pathology.
Safety analysis showed only 0.5% of patients (76 cases) had false-positive AI findings, and 92% of those were easily ruled out by radiologists. Of 28 false negatives, 89% were benign cysts, most smaller than 10mm. This balance of high sensitivity with manageable false-positive rates makes PANDA clinically viable.
PANDA was also validated on chest CT scans used routinely for lung cancer screening. Even though chest CT only partially images the upper abdomen and pancreas, PANDA successfully detected pancreatic lesions with meaningful accuracy.
This is particularly significant because chest CT for lung cancer screening is already a recommended public health intervention for high-risk individuals. If PANDA can be integrated into existing lung cancer screening programs, it could enable opportunistic pancreatic cancer screening at no additional cost or radiation exposure.
PANDA represents a major breakthrough: making early pancreatic cancer detection possible through scans that are already being done for other reasons. By repurposing routine non-contrast CT as a screening tool, PANDA could reach populations far larger than those covered by dedicated screening programs.
Future work will focus on prospective clinical trials to assess survival benefit from PANDA-assisted earlier detection, regulatory approvals, and integration into clinical information systems at scale. This study demonstrates that catching pancreatic cancer early is no longer the impossible task it once was.