Deep learning CT signature for predicting early liver metastases in pancreatic ductal adenocarcinoma

Abstract

Accurately predicting the risk of early liver metastases (ELM) and identifying patients who are most likely to benefit from neoadjuvant therapy (NAT) are critical for pancreatic ductal adenocarcinoma (PDAC). Here, we develop a Mamba-based predictive model that integrates imaging features from both the primary pancreatic tumor and the liver to assess the risk of ELM. The model is evaluated in a multi-institutional cohort of 1063 PDAC patients and demonstrates robust performance in predicting ELM (AUCs: 0.806-0.890). Besides, model-defined high-risk patients exhibit significantly shorter progression-free survival (PFS: HR = 1.93, p < 0.001) and overall survival (OS: HR = 1.89, p < 0.001). Notably, NAT confers significant OS benefits in model-defined high-risk patients (17.4 vs. 34.1 months; p < 0.001), even after propensity score matching (p = 0.004), while no survival benefit occurs in low-risk patients. Radiotranscriptomic analyses further reveal relative biological aggressiveness in the high-risk group. Overall, our proposed Mamba-based framework enables accurate prediction of ELM and may serve as a clinically actionable tool for identifying PDAC patients most likely to benefit from NAT. Predicting the risk of early liver micrometastases (ELM) is crucial for identifying patients who are most likely to benefit from neoadjuvant therapy (NAT) in pancreatic ductal adenocarcinoma (PDAC) yet most deep learning models focus mainly on primary tumours. Here, the authors show that the non-invasive Mamba based model is able to integrate multi-phase spatiotemporal CT imaging features from both the primary pancreatic tumour and liver regions to predict the risk of ELM in PDAC.