Multimodal Assessment of Pancreatic Cancer Resectability Using Deep Learning

Determining the resectability of pancreatic ductal adenocarcinoma (PDAC) is critical for surgical planning, yet it remains highly subjective and prone to inter-observer variability. Current clinical practice relies on assessing tumor-vessel interactions on CT scans, often supplemented by clinical markers. This study aims to develop a fully automated, multimodal deep learning tool that combines 3D CT imaging and structured clinical data to standardize the classification of PDAC into NCCN resectability categories (upfront resectable, borderline resectable, and locally advanced).

The researchers developed a multimodal framework using a Swin-UNETR backbone, which is well-suited for capturing both local and global 3D anatomical context. The model processes CT volumes to extract anatomical features while simultaneously using a multilayer perceptron to process 17 clinical variables (e.g., CA 19-9 levels, BMI, comorbidities). A key innovation is the use of a dynamic multitask objective: during training, the model performs auxiliary segmentation of the pancreas, tumor, and major vessels. This forces the encoder to learn 'vessel-aware' representations. Crucially, the segmentation decoder is only used during training, meaning the final model requires no manual or automated segmentation masks at inference time.

In a cohort of 159 patients, the model achieved an AUC of 0.86 and an accuracy of 0.85. External validation on an independent cohort of 52 patients yielded similar performance (AUC 0.86, accuracy 0.87), demonstrating the model's ability to generalize across different clinical settings. The study found that integrating clinical data with imaging features, guided by anatomical supervision, provides a more robust and reproducible assessment than unimodal or purely geometric approaches.

By automating the assessment of resectability, this tool could reduce the variability inherent in human judgment, potentially leading to more consistent treatment planning in multidisciplinary tumor boards. Because the model does not require segmentation masks at inference, it is easier to integrate into existing clinical workflows, offering a practical path toward objective, data-driven surgical decision support.