Introduction
Fatty pancreas is an increasingly recognized condition associated with metabolic abnormalities, but its diagnosis remains poorly standardized. In clinical practice, endoscopic ultrasound (EUS) is widely used to assess pancreatic parenchyma, yet its interpretation is mainly qualitative and operator-dependent, while magnetic resonance imaging (MRI) currently represents the most reliable non-invasive reference standard for pancreatic fat quantification.
Aims & Methods
This study aimed to assess the agreement between expert EUS-based diagnosis of fatty pancreas and MRI, and to develop deep learning models for automated fatty pancreas detection on EUS images. A retrospective single-centre observational study was conducted at San Raffaele Hospital including consecutive patients who underwent diagnostic EUS between January 2023 and December 2025 and had abdominal MRI within 12 months. Fatty pancreas was defined on MRI by pancreatic fat fraction >6%. Agreement between EUS and MRI was assessed with Cohen’s kappa. Secondary analyses included operator-level performance and subgroup analyses. For the AI endpoint, convolutional neural network models were developed using three approaches: all available EUS images per patient, a single representative image per patient, and a segmentation-based strategy combined with an ensemble model.
Results
Among 2324 EUS procedures, 314 patients had MRI available and were included in the primary analysis; 150 had fatty pancreas and 164 did not. Patients with fatty pancreas were older than controls (65.8 ± 11.0 vs 58.7 ± 14.2 years, p<0.001), and dyslipidaemia was more frequent (41% vs 25%, p<0.001). Agreement between EUS and MRI was only fair (Cohen’s κ=0.31), with sensitivity 39.9%, specificity 91.5%, and overall accuracy 66.6%. Marked inter-operator variability was observed, with κ ranging from 0.06 to 0.82. Concordance tended to be higher in patients with BMI >25 kg/m² (κ=0.40 vs 0.20, p=0.05) and was significantly lower in chronic pancreatitis (κ=-0.16 vs 0.34, p=0.03). For deep learning, the model using all available images achieved an AUC of 0.70. The single-image model achieved an AUC of 0.713, with high sensitivity (0.86–0.95) but modest specificity (0.38–0.50). The best performance was obtained with the segmentation-based ensemble model, which reached an AUC of 0.724, accuracy 69.8%, sensitivity 90%, specificity 52%, and balanced accuracy 0.71 in the independent test set.
Conclusion
EUS shows limited agreement with MRI for fatty pancreas diagnosis and substantial inter-operator variability, supporting the need for more objective and standardized diagnostic criteria. Deep learning applied to EUS images demonstrated moderate discriminatory ability, with the segmentation-based ensemble approach providing the most promising results. These findings support further development of AI-assisted EUS tools to improve reproducibility and diagnostic performance in fatty pancreas assessment.
