For pancreatic cancer patients, one of the most important decisions is whether surgery is possible. Resectability—the ability to surgically remove the tumor—depends heavily on how much the tumor has grown into or around nearby blood vessels. CT scans are used to make this determination, but different radiologists often reach different conclusions when reviewing the same scan.
This variability in assessment is not just an academic problem—it directly affects whether patients receive potentially curative surgery or are sent to palliative treatment instead. An AI model that could automatically and objectively quantify vascular involvement could standardize these decisions and potentially improve outcomes for patients at the borderline between operable and inoperable disease.
Researchers developed a semi-supervised machine learning segmentation model trained on 613 CT scans from 467 pancreatic tumor patients plus 50 control patients. The model was designed to automatically segment both the pancreatic tumor and five key blood vessels: the celiac trunk, hepatic artery, portal vein, superior mesenteric artery, and superior mesenteric vein.
A self-learning approach was used, where a 'teacher' model trained on a small set of manually labeled scans then generated labels for the remaining training data, creating a larger training set without requiring radiologists to manually annotate every case. After segmentation, the model measured vascular involvement by calculating the percentage of each vessel wall in contact with the tumor, then classified resectability as resectable, borderline resectable, or locally advanced.
On the test set of 60 patients (20 in each resectability category), the AI model agreed with expert radiologists on vascular involvement in 227 out of 300 vessel assessments (76%). While this shows room for improvement, it is comparable to the inter-observer variability between radiologists themselves.
For the overall resectability classification, the model agreed with radiologists in 85% of resectable cases, 80% of borderline resectable cases, and 75% of locally advanced cases. Borderline resectable is the most clinically consequential category—these are the patients most likely to benefit from or be harmed by surgery—and the AI performed well in this group.
Seven trained observers manually segmented tumors and vessels in 105 CT scans from patients with resectable, borderline resectable, and locally advanced disease, plus 50 control scans. This initial labeled set trained the teacher model. The teacher then generated pseudo-labels for additional unlabeled cases, which were reviewed and corrected where needed.
The model used a workflow consisting of three sequential automated steps: first segmenting the PDAC and surrounding vessels, then quantifying vascular involvement for each of the five relevant vessels, and finally classifying resectability using standardized Dutch Pancreatic Cancer Group criteria. This end-to-end automation required no radiologist input during inference.
The most significant clinical benefit of this AI tool would be standardizing resectability assessments across different hospitals and radiologists. Currently, whether a patient is offered surgery can depend significantly on which radiologist reviews their scan and which institution they attend. An objective AI measurement could make this decision more consistent and fair.
The tool could also serve as a decision support system, giving radiologists a quantitative second opinion. In borderline cases where radiologists are uncertain, having an objective vascular involvement measurement could help tip the balance toward or away from surgery with greater confidence.
This study demonstrates that AI can reliably automate the complex task of assessing vascular involvement and classifying tumor resectability in pancreatic cancer from CT scans. The agreement with expert radiologists was clinically meaningful and comparable to the variability observed between human readers.
The next steps involve prospective validation in clinical settings and testing whether AI-assisted decisions lead to better patient outcomes compared to standard radiologist assessment alone. If validated, this tool could significantly reduce unnecessary delays in surgical referrals and ensure more consistent care for pancreatic cancer patients.