Swin-UNETR deep learning model assesses pancreatic cancer resectability with 3 NCCN classes

Accurate determination of whether pancreatic cancer can be surgically removed is critical but suffers from high inter-expert variability. Researchers led by Vincent Ochs and colleagues from the University of Basel, University Hospital Basel, and Royal Free Hospital London introduced a fully automated multimodal deep learning framework that fuses 3D contrast-enhanced CT scans with 17 routinely collected clinical variables. The model uses a Swin-UNETR backbone to generate anatomy-aware representations by performing auxiliary segmentation of the pancreas, tumor, and major peripancreatic vessels. These features are combined with a compact clinical embedding and passed through a lightweight classification head to assign patients to one of three NCCN resectability categories: upfront resectable, borderline resectable, or locally advanced.

A key innovation is the dynamic multitask objective that adaptively weights segmentation and classification losses based on the current tumor Dice score. This encourages the model to learn features that are both anatomically informed and discriminative for resectability classification. The approach addresses two major pain points: the variability in human expert assessment and the need for automated, reproducible treatment planning. By leveraging both imaging and clinical data, the framework aims to provide consistent, evidence-based surgical decisions for pancreatic ductal adenocarcinoma (PDAC). The paper, available on arXiv (2607.13826), is submitted to the Computer Vision and Pattern Recognition domain and also spans AI, information retrieval, and machine learning.