Pancreatic ductal adenocarcinoma (PDAC) and chronic pancreatitis (CP) are two very different conditions that can look remarkably similar on imaging. Both can cause a solid mass in the pancreas, and both can present with pain, weight loss, and jaundice. Misdiagnosing one as the other leads to either delayed cancer surgery or unnecessary operations on benign disease.
Contrast-enhanced ultrasound (CEUS) is an imaging technique that uses tiny gas-filled microbubble contrast agents injected into the bloodstream. Unlike standard ultrasound, CEUS can show how blood flows through a lesion in real time, reflecting differences in vascularity between cancer and inflammation. However, interpreting CEUS images requires expertise that is not uniformly available.
This study investigated whether a deep learning radiomics (DLR) model trained on CEUS images could match or exceed the diagnostic accuracy of experienced radiologists in differentiating PDAC from CP, with the potential to standardize diagnosis across institutions with varying levels of expertise.
The study enrolled 558 patients across three hospitals: 291 with PDAC and 267 with CP, all confirmed by pathology. CEUS examinations were performed and representative image frames were extracted for analysis. This multi-center design was important for ensuring the model would generalize beyond the institution where it was trained.
The deep learning model used ResNet-50 as its backbone -- a 50-layer convolutional neural network that has proven highly effective at extracting hierarchical visual features from medical images. The model was trained on images from the primary hospital, with data from the other two hospitals used as external validation sets.
To understand what the model was learning, the researchers generated heatmaps (using Grad-CAM visualization) that overlaid color-coded attention regions onto the original CEUS images, showing which areas of each image most influenced the model's prediction.
A two-round reader study was conducted with five radiologists who first read cases without AI assistance, then re-read the same cases with the DLR model's output available. This design allowed direct measurement of how AI assistance changed diagnostic performance across readers with different experience levels.
The DLR model achieved an AUC of 0.986 on the training dataset and 0.978 on internal validation. Critically, performance remained strong on the two external validation sets from different hospitals, with AUCs of 0.967 and 0.953 -- a level of cross-site consistency that is difficult to achieve and indicates robust generalization.
In the reader study, AI assistance significantly improved the diagnostic performance of radiologists, particularly for less experienced readers. Sensitivity for PDAC detection increased when radiologists had access to the model output, meaning fewer cancer cases were missed when the AI was involved.
The model's heatmaps revealed biologically meaningful patterns: PDAC lesions showed enhancement primarily at the tumor periphery during the arterial phase -- consistent with the hypovascular nature of pancreatic adenocarcinoma. CP lesions showed more diffuse enhancement patterns, reflecting the different vascular architecture of inflamed tissue.
The Grad-CAM heatmaps provided visual evidence that the deep learning model was relying on diagnostically relevant features rather than spurious image artifacts. For PDAC cases, the model consistently highlighted the tumor margins and the interface between the mass and surrounding pancreatic tissue -- the same regions radiologists focus on.
For CP cases, the highlighted regions were more dispersed and corresponded to areas of inflammatory infiltration and fibrosis visible on CEUS. This divergence in attention patterns between the two conditions reflects genuine biological differences captured by the model.
The ability to generate these explanatory visualizations is crucial for clinical adoption. Radiologists are more willing to engage with and trust AI tools when they can see that the model is drawing on medically sound features, and heatmaps provide a language for that kind of collaborative human-AI reasoning.
One of the most practically significant findings was the differential benefit of AI assistance across radiologists with different experience levels. Less experienced radiologists showed larger improvements in sensitivity and specificity when using the AI as a second opinion, suggesting the model can act as a form of knowledge transfer or decision support.
This has important implications for healthcare equity: in hospitals without specialized pancreatic imaging expertise, a validated AI tool could help bring diagnostic quality closer to what is available at major academic centers. The consistent performance across three hospitals in this study provides early evidence that such cross-site deployment is feasible.
The two-round study design also highlights an important workflow consideration -- AI assistance changed decisions in a meaningful fraction of cases, and most of those changes were in the correct direction. This active correction of errors, rather than mere confirmation of correct diagnoses, is the mechanism through which AI tools add clinical value.
This study demonstrates that a ResNet-50-based DLR model can differentiate PDAC from chronic pancreatitis on CEUS images with very high accuracy that holds up across multiple independent hospital sites. Its combination of high performance, interpretability through heatmaps, and demonstrated benefit to radiologists positions it well for potential clinical use.
The multi-center design addresses one of the most common criticisms of AI in radiology -- that models trained at one institution fail when deployed elsewhere. The consistent AUCs above 0.95 at external sites suggest this model has learned generalizable disease features rather than site-specific imaging artifacts.
Prospective clinical trials, regulatory review, and integration into routine CEUS reporting workflows will be necessary before this tool can be widely deployed. However, the study provides a strong proof of concept that AI-assisted CEUS interpretation for PDAC diagnosis is both technically feasible and clinically meaningful.