Pancreatic neuroendocrine tumors (PNETs) and pancreatic ductal adenocarcinoma (PDAC) are both pancreatic cancers, but they differ fundamentally in prognosis, biology, and treatment. PNETs generally have a more favorable prognosis and respond to different therapies than PDAC. Correctly distinguishing them before treatment is therefore clinically critical - selecting the wrong treatment pathway can cause significant harm.
Endoscopic ultrasound (EUS) is the primary imaging modality for pancreatic lesion evaluation because of its high spatial resolution and ability to characterize lesion texture and vascularity in detail. However, accurate classification based on EUS images requires substantial expertise, and even experienced endosonographers face diagnostic uncertainty in a meaningful fraction of cases.
Deep learning models applied to EUS images can extract subtle quantitative features from images that are difficult or impossible to identify through visual inspection, potentially achieving diagnostic accuracy exceeding that of expert clinicians. An interpretable model that also explains its reasoning is particularly valuable in clinical settings where trust and accountability are essential.
The study enrolled 266 patients from a single institution: 151 with histologically confirmed pancreatic cancer (PDAC) and 115 with confirmed PNETs. EUS images were collected for each patient, and the dataset was split into training and test sets while maintaining the proportions of each tumor type.
ResNet18, a 18-layer convolutional neural network pretrained on ImageNet and fine-tuned for this task, was applied to extract 2048 deep learning features from each EUS image. This large feature set captures complex spatial patterns across multiple scales of image organization. LASSO regression was then applied to this feature set to select a compact, non-redundant subset of 27 features with the strongest discriminative value.
Two classification models were built: a Support Vector Machine (SVM) trained on the 27 DL features alone, and a Nomogram combining DL features with clinical variables. The nomogram format produces a visual tool that clinicians can use at the bedside by reading off scores from aligned scales for each variable and summing them to a total score that maps to a probability of PDAC versus PNET.
The SVM model using DL features achieved an AUC of 0.948 in training and 0.795 in the test set. The drop from training to test performance indicates some degree of overfitting to the training data, which is common with high-dimensional feature sets, and underscores the importance of reporting independent test set performance.
The combined Nomogram incorporating DL features and clinical variables achieved an AUC of 0.962 in training and 0.871 in testing, substantially improving over the DL-only model at both stages. This improvement confirms that clinical information (such as lesion size, hormone levels, or symptoms) complements imaging-derived features and should not be discarded in favor of pure image analysis.
DL_22 was identified as the most influential individual deep learning feature across both SHAP analysis and LASSO selection. While the specific visual texture pattern encoded in DL_22 requires further characterization, its consistent emergence across different analytical approaches validates it as a genuine imaging biomarker distinguishing these two tumor types.
Grad-CAM (Gradient-weighted Class Activation Mapping) was applied to visualize which regions of each EUS image the ResNet18 model focused on when making its classification. Heat maps overlaid on the original images showed that the model consistently attended to the internal texture of lesions and lesion boundaries rather than background tissue, confirming that the model is processing biologically relevant image regions.
SHAP (SHapley Additive exPlanations) values were computed for the 27 selected DL features to quantify each feature's contribution to the final classification decision. This provided a ranked list of feature importance that clinicians and researchers can use to understand which aspects of tumor imaging most distinguish PNETs from PDAC at the population level and for individual patients.
The combination of Grad-CAM spatial interpretability with SHAP feature importance creates a multi-level explanatory framework. A clinician can see both which part of the image the model found most informative (Grad-CAM) and which abstract extracted features contributed most to the diagnosis (SHAP), addressing two distinct dimensions of model transparency.
For community endoscopists with less experience in pancreatic EUS interpretation, a validated nomogram that achieves AUC 0.871 could serve as a decision support tool when facing diagnostically challenging lesions. The nomogram format - combining DL score with clinical inputs on a printed or digital tool - is accessible without specialized software infrastructure.
Accurate pre-treatment distinction between PNET and PDAC has direct implications for patient management. PDAC patients should receive systemic chemotherapy early and may be candidates for genetic testing (BRCA, PALB2, etc.), while PNET patients may be monitored conservatively for low-grade lesions or treated with somatostatin analogs, targeted agents such as everolimus or sunitinib, or peptide receptor radionuclide therapy (PRRT).
The interpretable nature of the model is particularly valuable for clinical acceptance, as physicians are more likely to trust and use AI tools they can interrogate rather than black-box systems. Regulatory pathways for AI-based medical imaging tools also increasingly require interpretability evidence, making the SHAP and Grad-CAM analyses practically important beyond academic interest.
This study developed an interpretable deep learning nomogram combining ResNet18-derived EUS features with clinical variables to distinguish PNETs from PDAC, achieving test set AUC of 0.871, substantially outperforming DL features alone (AUC 0.795). Grad-CAM and SHAP analyses provided mechanistic insight into model behavior.
The study's primary limitation is its single-institution design with 266 patients. External validation in diverse patient populations and EUS equipment configurations is necessary before clinical deployment. The test set performance improvement from 0.795 (DL-only) to 0.871 (nomogram) also suggests that the DL feature extraction could be further optimized with larger training datasets.
Future directions include multi-center validation, expansion to include other pancreatic lesion types such as mucinous cystic neoplasms and solid pseudopapillary tumors, and integration with EUS-guided fine needle biopsy results to create a multi-modal diagnostic pipeline. Development of a real-time implementation that processes EUS images during the procedure would maximize clinical utility.