Pancreatic tumors are among the hardest to detect and classify on CT scans. The pancreas is a small, deep organ surrounded by other structures, and tumors can appear subtle -- especially early-stage lesions. Radiologists must distinguish between multiple tumor types including pancreatic ductal adenocarcinoma (PDAC), intraductal papillary mucinous neoplasms (IPMN), and normal pancreas tissue, each requiring different management.
Current CT reading is time-consuming, requires expert radiologists, and is subject to inter-reader variability. In many healthcare settings, specialized pancreatic radiology expertise is not readily available, creating diagnostic gaps that can delay treatment or lead to misclassification.
This study developed a fully end-to-end deep learning (FEE-DL) system designed to automate the entire CT diagnostic workflow -- from screening individual images through locating the pancreas, segmenting the tumor, and finally classifying the lesion type -- without requiring manual intervention at any stage.
The FEE-DL system operates through four sequential stages. Stage 1 screens each CT slice to determine whether it contains diagnostically relevant information, filtering out non-pancreatic slices and reducing computational burden. Stage 2 uses a ResNet18 convolutional neural network to locate the pancreas within the relevant slices.
Stage 3 applies a U-Net32 architecture to segment the pancreas and any tumor region precisely, producing a pixel-level map of the tumor boundary. U-Net is widely used in medical image segmentation because of its ability to capture both fine detail and broader spatial context simultaneously.
Stage 4 uses a ResNet34 network to classify the segmented lesion as normal pancreas, PDAC, or IPMN. All four stages are fully automated and connected, meaning the output of each stage feeds directly into the next without human review at intermediate steps.
Training used 143,945 CT images from 319 patients, and testing was performed on a separate dataset of 107,036 images from 347 patients. This large-scale training and testing setup was designed to ensure the model could generalize across patient variation.
On the independent test set, FEE-DL achieved an AUC of 0.871, an F1 score of 88.5%, and an overall accuracy of 82.7%. PDAC was correctly identified in 87.6% of cases, while IPMN was detected at 100% sensitivity in the test cohort.
Importantly, the system processed each patient's complete CT study in an average of 18.6 seconds -- far faster than a human radiologist would take to review the same case. This speed, combined with consistent performance, suggests the system could serve as a rapid triage tool to flag suspicious cases for expert review.
The model's performance was evaluated separately for each diagnostic category, allowing a nuanced view of where the system succeeded and where errors occurred. PDAC detection accuracy was high but not perfect, and some misclassifications occurred between PDAC and normal pancreas, reflecting the subtle imaging appearances that challenge radiologists as well.
A key concern with deep learning in medicine is the black-box problem -- models may achieve high accuracy for reasons that are not clinically meaningful, making them difficult to trust and validate. To address this, the researchers generated saliency maps for each prediction.
Saliency maps highlight the regions of the CT image that most influenced the model's classification decision. In correctly classified PDAC cases, the saliency maps consistently pointed to the tumor region and adjacent pancreatic duct, which are the same features radiologists use when diagnosing PDAC. This alignment between AI attention and clinical reasoning increases confidence that the model is learning medically valid patterns.
When the model made errors, saliency analysis helped explain why -- for instance, pointing to regions of artifact or surrounding tissue rather than the tumor. This kind of interpretability tool is essential for building clinician trust and for identifying systematic failure modes that could be corrected in future model iterations.
Most prior AI studies in radiology addressed only one step in the diagnostic pipeline -- either segmentation or classification, but not both in a connected system. The end-to-end design of FEE-DL is significant because it eliminates the need for human review between stages and removes potential error accumulation from imperfect intermediate outputs being passed to the next step.
The large training and test datasets used in this study -- drawn from real clinical cases rather than carefully curated research datasets -- increase the likelihood that the model will maintain its performance when deployed in actual hospital settings. Models trained on clinical data tend to be more robust to the variability inherent in routine practice.
There are practical deployment considerations, including integration with hospital picture archiving and communication systems (PACS), regulatory approval, and the question of how to best incorporate AI outputs into clinical workflows without creating alert fatigue or undermining radiologist judgment.
FEE-DL demonstrates that a fully automated deep learning system can handle the complete CT diagnostic workflow for pancreatic tumors with accuracy comparable to expert radiologists. Its speed -- under 20 seconds per patient -- and end-to-end design make it a strong candidate for high-throughput screening applications.
With pancreatic cancer often detected only at late stages when surgery is no longer possible, any technology that can improve early detection rates could have a substantial impact on survival. Automated AI triage could flag suspicious studies for urgent expert review, potentially shortening the time between imaging and diagnosis.
Future development should include multi-center validation, prospective clinical trials, and investigation of whether the system can also detect earlier-stage tumors where current clinical practice most frequently misses disease. Expanding the system to classify additional pancreatic lesion types would further enhance its clinical utility.