Introduction

The use of contrast-enhanced endoscopic ultrasound (CE-EUS) for characterizing solid pancreatic masses has improved diagnostic capabilities. We aimed to evaluate the performance of an artificial intelligence (AI)-powered system for differential diagnosis of pancreatic adenocarcinoma using CE-EUS.

Aims & Methods

The methodology integrates advanced imaging analysis and machine learning (ML) to develop an automated framework for pancreatic tumor characterization using EUS. A convolutional neural network (CNN) is first trained on labeled non-contrast EUS images to detect and segment the pancreas and tumor regions accurately. Once trained, the CNN is applied to new cases, identifying regions of interest (ROIs) in the non-contrast images, which are subsequently mapped onto the corresponding contrast-enhanced images for further analysis. Time-Intensity Curve (TIC) analysis was performed on the mapped ROIs, extracting a comprehensive set of quantitative parameters describing contrast enhancement dynamics. These parameters include temporal metrics (e.g., time to peak, rise time), intensity measures (e.g., peak intensity, wash-in intensity), slopes (e.g., wash-in and wash-out rates), area-based metrics (e.g., area under the curve), and perfusion-related characteristics (e.g., perfusion index, blood flow). The extracted TIC parameters are then used to train a feedforward neural network (FNN), which learns to differentiate diagnostic categories based on the perfusion characteristics of the lesions. In the final integrated model, the CNN detects pancreas and tumor regions, maps them to contrast-enhanced images, computes TIC parameters, and feeds these into the FNN to provide automated diagnostic predictions.

Results

Retrospective analysis was performed on a prospectively collected dataset of 120 CE-EUS videos obtained from patients with solid pancreatic masses in 3 tertiary referral centers. The dataset included 88 pancreatic adenocarcinomas, 1 lymphoma, 4 metastatic lesions, 10 neuroendocrine neoplasms (G1/G2), 15 cases of chronic pancreatitis, and 2 solid pseudopapillary neoplasms (SPNs). Diagnostic performance was evaluated by calculating sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy. Based on CNN analysis, both precision and recall were also calculated. The AI-powered CE-EUS system achieved high diagnostic performance, with an accuracy of over 95% for the diagnosis of pancreatic adenocarcinoma. Precision and recall were also excellent for differentiating adenocarcinoma from other pancreatic lesions based on CNN analysis.

Conclusion

This novel AI-based approach using CE-EUS reliably distinguished pancreatic adenocarcinoma from other solid pancreatic masses. Real-time analysis of time-intensity curve parameters enhances diagnostic accuracy and offers a promising adjunct to traditional EUS methods. Further prospective validation is warranted to confirm these findings and explore clinical applicability.

References

1. Săftoiu A, Vilmann P, Bhutani MS, Hassan H, Hawes R, Dietrich CF, et al. Enhanced endoscopic ultrasound imaging for pancreatic lesions: A review.
World J Gastrointest Endosc. 2022 Jul 16;14(7):369–84. doi:10.4253/wjge.v14.i7.369 PMID: 35967758; PMCID: PMC9367228.
2. Liang T, Zhang H, Wang Y, Sun Y, Ma K, Liu W, et al. Artificial intelligence-assisted endoscopic ultrasound for detection of pancreatic lesions: A systematic review and meta-analysis. World J Gastroenterol. 2023 Feb 21;29(7):1150–1164. doi:10.3748/wjg.v29.i7.1150 PMID: 36849883; PMCID: PMC10720860.