Nicoleta Podină

Background/Objectives: Contrast-enhanced harmonic imaging endoscopic ultrasound (CHI-EUS) is a valuable tool for characterizing solid pancreatic tumors. However, interobserver variability remains a significant limitation in clinical interpretation. Artificial intelligence (AI) may offer objective, reproducible assessments, potentially enhancing diagnostic performance. This study compared the diagnostic accuracy and interobserver agreement of nine physicians with varying CHI-EUS experience levels vs. a dedicated AI system and a general-purpose large language model (ChatGPT) on the same 118 histologically confirmed cases. Methods: We conducted a prospective, multicenter, observer-blinded study involving 118 CHI-EUS video cases of histologically confirmed (EUS-FNB) focal pancreatic masses from three tertiary care centers in Romania. Nine readers were stratified into three groups: trainees (10 years). All readers and two AI models received standardized, anonymized 2 min CHI-EUS video clips. A dedicated AI system used a convolutional neural network (CNN) for lesion segmentation and time–intensity curve (TIC) extraction, followed by a feedforward neural network (FNN) for classification. ChatGPT was separately evaluated on the same videos. Diagnostic metrics (accuracy, sensitivity, specificity, positive predictive value [PPV], negative predictive value [NPV], and AUROC) were calculated. Interobserver agreement was assessed using Fleiss’ and Cohen’s kappa statistics. Results: The dedicated AI system achieved an overall accuracy of 95.8% (sensitivity 96.6%; specificity 94.1%) in diagnosing pancreatic adenocarcinoma. Expert readers had a mean accuracy of 78.8% (sensitivity 86%, specificity 61%, and AUROC 0.74), intermediates 80.8% (sensitivity 83%, specificity 75%, and AUROC 0.84), and trainees had a mean accuracy of 67.2% (sensitivity 70%, specificity 60%, and AUROC 0.67). For the most-likely-diagnosis parameter, interobserver agreement was similar between intermediates (Fleiss’ κ = 0.407) and experts (κ = 0.389), while trainees showed lower agreement (κ = 0.203). ChatGPT correctly classified only 14.1% of PDAC cases. Conclusions: A specialized AI model for CHI-EUS video analysis can achieve expert-level performance and reduce diagnostic variability across experience levels. Integration of dedicated AI systems into CHI-EUS interpretation may enhance accuracy and serve as a valuable decision support tool in clinical and training settings.

Background: EUS-guided fine-needle biopsy is the procedure of choice for the diagnosis of pancreatic ductal adenocarcinoma (PDAC). Nevertheless, the samples obtained are small and require expertise in pathology, whereas the diagnosis is difficult in view of the scarcity of malignant cells and the important desmoplastic reaction of these tumors. With the help of artificial intelligence, the deep learning architectures produce a fast, accurate, and automated approach for PDAC image segmentation based on whole-slide imaging. Given the effectiveness of U-Net in semantic segmentation, numerous variants and improvements have emerged, specifically for whole-slide imaging segmentation. Methods: In this study, a comparison of 7 U-Net architecture variants was performed on 2 different datasets of EUS-guided fine-needle biopsy samples from 2 medical centers (31 and 33 whole-slide images, respectively) with different parameters and acquisition tools. The U-Net architecture variants evaluated included some that had not been previously explored for PDAC whole-slide image segmentation. The evaluation of their performance involved calculating accuracy through the mean Dice coefficient and mean intersection over union (IoU). Results: The highest segmentation accuracies were obtained using Inception U-Net architecture for both datasets. PDAC tissue was segmented with the overall average Dice coefficient of 97.82% and IoU of 0.87 for Dataset 1, respectively, overall average Dice coefficient of 95.70%, and IoU of 0.79 for Dataset 2. Also, we considered the external testing of the trained segmentation models by performing the cross evaluations between the 2 datasets. The Inception U-Net model trained on Train Dataset 1 performed with the overall average Dice coefficient of 93.12% and IoU of 0.74 on Test Dataset 2. The Inception U-Net model trained on Train Dataset 2 performed with the overall average Dice coefficient of 92.09% and IoU of 0.81 on Test Dataset 1. Conclusions: The findings of this study demonstrated the feasibility of utilizing artificial intelligence for assessing PDAC segmentation in whole-slide imaging, supported by promising scores.