Surgically removing a pancreatic tumor—a procedure called a pancreaticoduodenectomy or Whipple operation—is one of the most complex operations performed in abdominal surgery. The pancreas sits in a tight anatomical space surrounded by major blood vessels, bile ducts, and the duodenum, and anatomic variations between patients make every operation unique.
A critical part of surgical planning is understanding exactly how the tumor relates to nearby vessels. Some tumors grow around or into arteries and veins in ways that make complete removal impossible. Surgeons need precise 3D maps of this anatomy before making an incision, but creating these maps manually from CT scan slices is time-consuming and imprecise.
3D surgical simulation has already transformed liver surgery planning. This Japanese study evaluated whether a deep learning-powered system called SYNAPSE VINCENT could do the same for pancreatic surgery, automatically generating 3D models of the pancreas, pancreatic duct, and surrounding vasculature from standard CT images.
Researchers at Saitama Cancer Center enrolled 100 pancreatic cancer patients and 100 non-cancer control patients. All underwent CT scanning, and the SYNAPSE VINCENT deep learning system was used to automatically segment and reconstruct the pancreatic tissue, pancreatic duct, portal vein, arteries, and surrounding venous structures in three dimensions.
The accuracy of these AI-generated models was measured using the Dice Coefficient (DC)—a standard metric comparing how well the AI's segmentation overlaps with manual expert segmentation. DC values range from 0 (no overlap) to 1 (perfect agreement), with values above 0.8 generally considered clinically acceptable.
The team also analyzed whether tumor characteristics—size, location, and cancer stage—affected the AI's ability to accurately model the pancreatic tissue, and whether duct diameter correlated with model accuracy.
The deep learning system achieved Dice Coefficients of 0.83 for pancreatic tissue in cancer patients and 0.86 in non-cancer patients, both well within clinically acceptable ranges. For pancreatic duct reconstruction, the AI achieved DC of 0.84 in cancer patients and 0.77 in non-cancer patients.
For the critical vascular structures, the AI performed well on arteries (DC 0.89 in cancer patients) and portal veins (DC 0.89), slightly lower on other veins (DC 0.85), and acceptably on smaller venous branches (DC 0.82). These are the vessels surgeons most need to visualize when assessing resectability.
Notably, the accuracy was not affected by tumor size, location within the pancreas (head, body, or tail), or cancer stage. This is an important finding because it suggests the model performs reliably even for advanced or unusually located tumors. The AI-measured pancreatic duct diameter also correlated positively with manual measurement (r=0.61), confirming practical utility.
Before operating, surgeons can use these AI-generated 3D models to visualize the exact path they will need to take and identify vessel anatomy that might not be obvious from 2D CT images. This is particularly valuable for identifying anatomic variations—such as unusual arterial branching patterns—that could lead to dangerous intraoperative bleeding if unexpected.
The models also help surgeons and patients have more meaningful conversations about risk. A 3D visualization showing a tumor wrapped around a major artery is far more communicable than a radiologist's text report, enabling more informed consent and shared decision-making about whether to attempt surgery.
The system is semi-automated, meaning a technician can generate a 3D model in far less time than manual segmentation would require. As the algorithm improves and becomes fully automated, this technology could become a standard part of the preoperative workup for every pancreatic cancer patient being considered for surgery.
This study demonstrates that deep learning can reliably reconstruct the complex anatomy relevant to pancreatic surgery from routine CT scans. The consistent accuracy across tumor sizes and stages, and across both cancer and non-cancer patients, supports the broader clinical applicability of this approach.
Future improvements could include incorporating intraoperative imaging to update the model in real time during surgery, and training the AI on larger, more diverse datasets to capture an even wider range of anatomic variations. Integration with surgical navigation systems represents the next frontier.