A Causality-Informed Graph Intervention Model for Pancreatic Cancer Early Diagnosis

Abstract

Pancreatic cancer is a highly fatal cancer type. Patients are typically in an advanced stage at their first diagnosis, mainly due to the absence of distinctive early stage symptoms and lack of effective early diagnostic methods. In this work, we propose an automated method for pancreatic cancer diagnosis using noncontrast computed tomography (CT), taking advantage of its widespread availability in clinic. Currently, a primary challenge limiting the clinical value of intelligent systems is low generalization, i.e., the difficulty of achieving stable performance across datasets from different medical sources. To address this challenge, a novel causality-informed graph intervention model is developed based on a multi-instance-learning framework integrated with graph neural network (GNN) for the extraction of local discriminative features. Within this model, we develop a graph causal intervention scheme with three levels of intervention for graph nodes, structures, and representations. This scheme systematically suppresses noncausal factors and thus lead to generalizable predictions. Specifically, first, a target node perturbation strategy is designed to capture target-region features. Second, a causal-structure separation module is developed to automatically identify the causal graph structures for obtaining stable representations of whole target regions. Third, a graph-level feature consistency mechanism is proposed to extract invariant features. Comprehensive experiments on large-scale datasets validated the promising early diagnosis performance of our proposed model. The model generalizability was confirmed on three independent datasets, where the classification accuracy reached 86.3%, 80.4%, and 82.2%, respectively. Overall, we provide a valuable potential tool for pancreatic cancer screening and early diagnosis.