Early and Prediagnostic Detection of Pancreatic Cancer from Computed Tomography

Pancreatic ductal adenocarcinoma (PDAC), one of the deadliest solid malignancies, is often detected at a late and inoperable stage. Retrospective reviews of prediagnostic CT scans, when conducted by expert radiologists aware that the patient later developed PDAC, frequently reveal lesions that were previously overlooked. To help detecting these lesions earlier, we developed an automated system named ePAI (early Pancreatic cancer detection with Artificial Intelligence). It was trained on data from 1,598 patients from a single medical center. In the internal test involving 1,009 patients, ePAI achieved an area under the receiver operating characteristic curve (AUC) of 0.939-0.999, a sensitivity of 95.3%, and a specificity of 98.7% for detecting small PDAC less than 2 cm in diameter, precisely localizing PDAC as small as 2 mm. In an external test involving 7,158 patients across 6 centers, ePAI achieved an AUC of 0.918-0.945, a sensitivity of 91.5%, and a specificity of 88.0%, precisely localizing PDAC as small as 5 mm. Importantly, ePAI detected PDACs on prediagnostic CT scans obtained 3 to 36 months before clinical diagnosis that had originally been overlooked by radiologists. It successfully detected and localized PDACs in 75 of 159 patients, with a median lead time of 347 days before clinical diagnosis. Our multi-reader study showed that ePAI significantly outperformed 30 board-certified radiologists by 50.3% (P < 0.05) in sensitivity while maintaining a comparable specificity of 95.4% in detecting PDACs early and prediagnostic. These findings suggest its potential of ePAI as an assistive tool to improve early detection of pancreatic cancer.

💡 Research Summary

Pancreatic ductal adenocarcinoma (PDAC) remains one of the deadliest solid tumors, with a five‑year survival of only about 13 % because most cases are diagnosed after the disease has already spread beyond the pancreas. Early detection dramatically improves outcomes, raising five‑year survival to 40‑45 % when the tumor is still confined to the pancreas. Computed tomography (CT) is performed at massive scale worldwide—roughly 300 million examinations per year, many of them contrast‑enhanced abdominal scans that already include the pancreas—making it an attractive substrate for opportunistic cancer screening. However, conventional radiologist interpretation often misses subtle early changes, especially in non‑pancreatic CT studies, and retrospective reviews of pre‑diagnostic scans have shown low sensitivity (≈30 %).

To address this gap, the authors developed an automated, open‑source AI system called ePAI (early Pancreatic cancer detection with Artificial Intelligence). ePAI follows a three‑stage cascade: (1) anatomical segmentation of the pancreas, surrounding vessels, and dilated ducts using nnU‑Net; (2) detection and localization of all potential pancreatic lesions, trained on both real lesions and a large set of synthetically generated lesions created by generative AI to boost sensitivity for very small tumors; (3) lesion‑level classification into PDAC, non‑PDAC, or normal using a combination of radiomic, texture, shape, location, and deep‑learning features.

Training data comprised 1,598 contrast‑enhanced abdominal CT scans from Johns Hopkins Hospital, with ground‑truth labels confirmed by surgical pathology (for lesions) or two‑year follow‑up (for normals). Pixel‑wise annotations of pancreas and lesions were provided by radiologists.

Internal validation used 1,009 patients (279 PDAC, 427 non‑PDAC, 303 normals). ePAI achieved an AUC of 0.985 for all‑size PDAC detection, with sensitivity 97.1 % and specificity 98.7 %. For tumors ≤2 cm the sensitivity was 95.3 % and the system could localize lesions as small as 2 mm. Lesion‑level localization accuracy was 94.6 % overall and 88.2 % for small tumors.

External multicenter validation involved six independent cohorts (total 7,158 patients: 610 small PDAC, 2,529 large PDAC, 4,019 normals) across North America, Europe, and Asia. ePAI maintained strong performance: AUC 0.971, sensitivity 97.0 % (overall) and 91.5 % for ≤2 cm tumors, specificity 88.0 %. Stage‑wise sensitivities were 90.2 % (T1), 96.4 % (T2), and 98.4 % (T3‑T4). Localization accuracy in the external set was 88.7 % overall and 79.7 % for small tumors.

Prediagnostic detection was evaluated on 159 patients who had both a pre‑diagnostic CT (3–36 months before clinical diagnosis) and a diagnostic CT. Radiologists missed all lesions on the pre‑diagnostic scans, whereas ePAI correctly identified PDAC in 75 patients (47 % detection rate) with a median lead time of 347 days. When detection occurred, localization matched the eventual tumor location in 75 % of cases, and secondary morphological cues such as pancreatic duct dilation were identified in 85 % of patients.

A multi‑reader study compared ePAI against 30 board‑certified radiologists on the same pre‑diagnostic dataset. ePAI’s sensitivity was 50.3 % higher than the human readers (p < 0.05) while maintaining comparable specificity (95.4 %).

Key strengths of the work include: (1) pixel‑level interpretability through explicit segmentation masks; (2) use of synthetic lesions to augment training and improve detection of sub‑centimeter disease; (3) open‑source release of code, model weights, and detailed annotation protocols, facilitating reproducibility and independent benchmarking.

Limitations are acknowledged: the training cohort originates from a single institution, raising questions about performance on rare histologic subtypes or non‑contrast CT protocols; the system was not fine‑tuned on pre‑diagnostic scans, so false‑positive rates in routine screening settings need further assessment; and prospective clinical impact, cost‑effectiveness, and workflow integration remain to be demonstrated.

In conclusion, ePAI demonstrates that a carefully engineered AI pipeline can reliably detect and localize early‑stage pancreatic cancer on both diagnostic and pre‑diagnostic CT scans across diverse populations. By identifying tumors on average nearly a year before they become apparent to radiologists, ePAI holds promise as an assistive tool that could shift PDAC diagnosis toward earlier, potentially curable stages, ultimately improving patient outcomes.