Introduction
Pancreatic Ductal Adenocarcinoma (PDAC) remains one of the most lethal malignancies, with surgical resection offering the only potential for a long-term cure. However, at the time of diagnosis, only approximately 20% of patients are candidates for surgery. The primary determinant of resectability is the degree of tumor involvement with adjacent major vascular structures, such as the superior mesenteric artery and vein. Currently, this assessment relies on manual radiological evaluation of "degrees of contact" on CT scans—a process characterized by high inter-observer variability and subjectivity. This paper presents an automated deep learning workflow designed to provide objective, quantifiable, and reliable assessments of tumor-vessel involvement to aid clinical decision-making.
Methodology: A Multi-Stage Pipeline
The researchers developed a comprehensive pipeline to automate the assessment of resectability. The process begins with the extraction of a Region of Interest (ROI). Given the complexity of abdominal CT scans, the system first segments the pancreas to define a localized volume for further analysis. This step ensures that subsequent models focus on the most relevant anatomical structures.
Following ROI extraction, the pipeline employs multi-structure segmentation using three distinct 3D architectures: nnU-Net, a custom 3D U-Net, and a Probabilistic 3D U-Net. These models are tasked with segmenting the tumor, veins, and arteries. A key innovation in this work is the introduction of the Overlap Loss (OLL). Traditional segmentation metrics like the Dice coefficient often fail to capture the nuances of the thin interface where a tumor touches a vessel. The OLL specifically penalizes errors in these contact zones, significantly improving the model's ability to detect subtle involvement.
Quantifying Involvement and Uncertainty
Once segmentation is complete, the system automatically computes the contact angle. By identifying the vessel's centroid and the arc of tumor contact on each slice, the tool provides a precise measurement of involvement. Recognizing that tumor boundaries are often ill-defined on CT scans, the authors integrated uncertainty estimation. Using model ensembles and probabilistic modeling, the system generates confidence intervals (e.g., ±1σ or ±2σ). This allows clinicians to see not just a single predicted angle, but a range of possible involvements based on model confidence.
Results and Discussion
The study utilized a dataset of 195 CT scans from 99 patients, meticulously annotated by expert radiologists. The models demonstrated high accuracy in segmenting vascular structures, with Dice scores of 0.88 for veins and 0.86 for arteries. Tumor segmentation proved more challenging (Dice 0.66) due to the inherent low contrast of PDAC.
Crucially, the system achieved 88.2% sensitivity and 85.7% specificity in detecting tumor-vessel involvement. The Probabilistic 3D U-Net showed the strongest correlation with ground-truth venous contact angles ($R^2 = 0.60$). The impact of the Overlap Loss was particularly notable, nearly doubling the Dice score for the critical tumor-vein interface compared to baseline methods.
Conclusion
This automated approach addresses the critical need for objective resectability assessment in pancreatic cancer. By providing precise contact angles and visualizing uncertainty, the tool serves as a robust decision support system for surgeons. Future work will focus on expanding the dataset—particularly for rare arterial involvement cases—and incorporating broader clinical context to further refine the predictive accuracy of the model.
Summary updated