Bi-parametric magnetic resonance imaging (bpMRI) is widely used for non-invasive diagnosis of clinically significant prostate cancer (csPCa). Although the Prostate Imaging Reporting and Data System (PI-RADS) standardizes interpretation through structured diagnostic criteria, clinical assessment remains highly operator-dependent. Vision–language models (VLMs), such as CLIP, offer a promising automated solution by leveraging the imaging-language coupling inherent in PI-RADS. However, existing VLM adaptation methods generally treat medical diagnosis as a generic image-to-label matching task. This conventional “label-driven” paradigm relies on superficial semantic correlations, fundamentally failing to capture the structured reasoning, anatomical context, and sequence prioritization defined by clinical guidelines like PI-RADS. To address this fundamental limitation, we propose KAMP-Net, a multi-modal prompt-tuning network that shifts the paradigm toward “knowledge-driven” learning. Our core algorithmic novelty lies in mathematically translating structured PI-RADS criteria directly into the network’s computational graph. Specifically, rather than employing generic feature fusion and global alignment, KAMP-Net utilizes zone-aware modality routing to dynamically gate heterogeneous bpMRI sequences, and enforces a hierarchical vision–language alignment that structurally mirrors a radiologist’s bottom-up diagnostic workflow. Experiments on a multi-center cohort comprising two public and one private datasets, plus an external private dataset, show that KAMP-Net delivers competitive performance in both patient-level diagnosis and lesion-level localization. Beyond standalone performance, we evaluated its clinical utility in an AI-assisted workflow. Notably, for challenging equivocal cases (PI-RADS = 3), this human-in-the-loop strategy successfully reduces false positives without risking missed diagnoses, demonstrating its strong potential as a reliable decision-support tool. Code is available at: https://github.com/IsYuchenYuan/KAMP-Net.
Patient-specific haemodynamic assessment is essential for clarifying vascular pathophysiology and guiding clinical decision-making. While computational fluid dynamics offers high-fidelity flow quantification, its routine clinical use remains limited by the time-consuming iterative solution process required for patient-specific simulations. To tackle these issues, we developed a geometry-guided graph encoding network (G3e-Net), a CFD-supervised image-derived prediction framework that learns anatomically informed representations from CTA-derived vascular geometries. G3e-Net coupled centerline-guided spectral encoding to resolve global topology with multi-scale curvature descriptors capturing local boundary nuances. An attention mechanism then combined these different spatial representations, allowing the network to dynamically balance global connectivity with local geometry in complicated vascular structures. By learning a direct mapping from CTA-derived point clouds to three-dimensional haemodynamic fields, G3e-Net enables rapid functional estimation without requiring volumetric meshing or iterative numerical optimization during inference. Extensive validation on thoracic aortas and carotid arteries from multiple hospitals confirmed that G3e-Net outperformed state-of-the-art baselines, reducing prediction errors for velocity, pressure, and WSS by 10-20%. Notably, inference on high-resolution aortic geometries comprising approximately 200,000 nodes was completed in less than 4 seconds, providing substantial acceleration compared with conventional CFD. These results support the potential of anatomically informed graph representation learning for rapid image-derived haemodynamic analysis and patient-specific functional vascular assessment in thoracic aortic and carotid artery applications.
