What question did this study set out to answer?

The aim is to evaluate pancreatic cancer operability using automated deep learning and multimodal data integration.

March 29, 2026

SRS388 - Pancreatic resectability evaluation through deep imaging with computed tomography in pancreatic cancer

VOVincent Ochs CKChristoph KümmerliNIHR Southampton Biomedical Research Centre JWJulia WollebYale University

Key Points

The aim is to evaluate pancreatic cancer operability using automated deep learning and multimodal data integration.
Developed a pipeline integrating contrast-enhanced CT and clinical data for operability assessment.
Processed CT scans with a SwinUNETR architecture for imaging feature extraction and segmentation.
Utilized a multilayer perceptron for classification based on embedded imaging and clinical variables.
Conducted model training with multi-task loss functions and evaluated using stratified 5-fold cross-validation.
Among 159 patients, 53.5% were classified as resectable.
Achieved a mean Dice score of 75.2% across 16 classes for segmentation accuracy.
Operability prediction accuracy reached 85.1%, outperforming previous methods.

Abstract

Abstract Introduction Pancreatic ductal adenocarcinoma (PDAC) has the highest mortality of solid malignancies (Sung H, CA Cancer J Clin). Surgical resection is the only curative option (NCCN Guidelines, 2019), yet accurate determination of resectability is difficult. Defined by vascular involvement, assessment suffers from low inter-observer agreement. Automated deep learning may provide standardized, reproducible support for operability assessment. Methods We developed an end-to-end multimodal pipeline to predict PDAC operability (resectable, borderline, irresectable) by integrating contrast-enhanced CT and clinical data. CT scans were processed using a SwinUNETR backbone. The encoder produced imaging features, while the decoder generated 16-class segmentations of pancreas, tumor, and vessels. Seventeen clinical variables underwent projection into an embedding. Imaging and tabular embeddings were concatenated and classified via a multilayer perceptron. Training used a multi-task loss combining segmentation (Dice, weighted cross-entropy, focal) and classification (weighted cross-entropy) with adaptive weighting. Evaluation used stratified 5-fold cross-validation and grid search optimization. Results Among 159 patients, 85 (53.5%) were resectable, 47 (29.5%) borderline, and 27 (17%) irresectable. Cross-validation yielded a mean Dice of 75.2% across 16 classes, including 70.4% for tumor, 69.7% for the superior mesenteric artery, and 65.1% for the vein. Operability prediction reached 85.1% accuracy, exceeding prior methods (70–88%) on smaller datasets. Conclusions Our model shows feasibility for automated PDAC operability prediction. By leveraging segmentation as an auxiliary task, it learns anatomically relevant features enhancing classification. Unlike segmentation-only approaches (Viviers, Bereska), our pipeline fuses imaging and clinical data. Future work includes multi-institutional validation and a prospective clinical trial via Pancreasgroup.org.

Ask AI

Helpful

Bookmark

View Full Paper