SeekFlow: Synergizing Radiology and Pathology Foundation Models for Precision Oncology via Knowledge-Guided Evidence Flow
This repository contains the official code for SeekFlow: Synergizing Radiology and Pathology Foundation Models for Precision Oncology via Knowledge-Guided Evidence Flow, accepted by ECCV 2026.
SeekFlow is a multimodal learning framework for precision oncology that synthesizes macro-scale radiology evidence and micro-scale pathology evidence from heterogeneous foundation models. Instead of relying on static feature fusion, SeekFlow reformulates multimodal integration as a knowledge-guided evidence flow problem. It transports radiology and pathology tokens toward a shared evidence manifold structured by clinical prototypes, preserving feature topology while disentangling shared and modality-specific evidence. The framework further introduces an uncertainty-aware evidential synthesis module to aggregate transported cross-modal evidence for reliable diagnosis and prognosis. Experiments across TCGA Glioma, Gastric Cancer, and Chondrosarcoma datasets show that SeekFlow achieves state-of-the-art performance on both diagnostic and prognostic tasks.