Problem

Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal cancer often diagnosed at an advanced stage when treatment options are limited. While computed tomography (CT) scans are widely used and can potentially reveal early signs of PDAC, human radiologists frequently overlook subtle lesions, especially when scans are performed for other reasons. Existing AI methods for PDAC detection often suffer from limitations such as small evaluation cohorts, poor performance on small tumors, lack of generalizability across diverse populations, and absence of precise lesion localization. This highlights a critical need for an advanced computational tool to assist in the early and prediagnostic detection of PDAC.

Method

The researchers developed ePAI, an open-source, automated system for detecting and localizing PDAC from contrast-enhanced CT scans. ePAI employs a three-stage cascade architecture: Stage 1 performs anatomical segmentation of the pancreas and surrounding structures; Stage 2 detects and localizes potential pancreatic lesions using an nnU-Net model fine-tuned on both real and synthetic lesions to enhance sensitivity for small tumors; and Stage 3 classifies these detected lesions as PDAC, non-PDAC, or normal based on various features. The system was trained on 1,598 patient CT scans from Johns Hopkins Hospital, with ground truth confirmed by surgical pathology or two-year follow-up. Evaluation was conducted on a diverse internal test cohort (1,009 patients) and a large external multicenter cohort (7,158 patients across 6 centers) to assess generalizability. Crucially, ePAI was also tested on prediagnostic CT scans obtained 3 to 36 months before clinical diagnosis, where radiologists had originally missed the cancer. A multi-reader study involving 30 radiologists further compared ePAI's performance to human interpretation and assessed the impact of AI assistance.

Results

ePAI demonstrated strong performance across all evaluations. In the internal test, it achieved an AUC of 0.985, a sensitivity of 97.1%, and a specificity of 98.7% for all PDACs, with 95.3% sensitivity for small PDACs (≤2 cm), detecting lesions as small as 2 mm. External validation across six centers yielded an AUC of 0.971, sensitivity of 97.0%, and specificity of 88.0% for all PDACs, maintaining 91.5% sensitivity for small PDACs. Significantly, on prediagnostic CT scans, ePAI detected PDAC in 75 out of 159 patients, providing a median lead time of 347 days before radiologists made a clinical diagnosis. In the multi-reader study, ePAI significantly outperformed 30 radiologists by 50.3% in sensitivity while maintaining a comparable specificity of 95.4% for early and prediagnostic detection. When assisted by ePAI, radiologists' sensitivity improved by 19.6% and 6.9% in diagnostic and prediagnostic scans, respectively.

Implications

The ePAI system shows significant promise as an assistive tool for improving early and prediagnostic detection of pancreatic cancer. Its robust performance across diverse external cohorts suggests broad clinical deployability. The strong localization capabilities can reduce radiologists' search burden and mitigate perceptual errors, enabling targeted follow-up. The ability to detect PDAC months or years before clinical diagnosis highlights its potential to facilitate earlier intervention and improve patient outcomes, especially for curable early-stage cancers. Future work should integrate ePAI with clinical risk models to optimize its use in low-prevalence settings and conduct prospective clinical validations. Limitations include incomplete clinical information for some external sites, lack of interpretation time measurement in the multi-reader study, and performance variations across centers due to imaging heterogeneity.