STG: Spatiotemporal Graph Neural Network with Fusion and Spatiotemporal Decoupling Learning for Prognostic Prediction of Colorectal Cancer Liver Metastasis

Yiran Zhu, Wei Yang, Yan Su, Zesheng Li, Chengchang Pan, Honggang QiarXiv 2025 · DOI · arXiv

Introduction

Predicting postoperative outcomes for colorectal cancer liver metastasis (CRLM) is a significant clinical challenge. While surgical intervention can remove visible tumors, recurrence rates remain high, and individual disease trajectories vary widely. Traditional prognostic tools often rely on static "snapshots" or simplistic risk scores, failing to capture the dynamic nature of cancer progression.

This paper introduces DyPro, a multimodal deep learning framework designed to simulate future disease evolution in a latent space. By integrating preoperative CT imaging and clinical data into a dynamic graph-based model, DyPro provides more accurate predictions for Disease-Free Survival (DFS) and Overall Survival (OS).

Problem Statement

Existing prognostic models for CRLM face three primary limitations:

  1. Static Analysis: Most models utilize a single preoperative scan, ignoring the temporal evolution of the disease.
  2. Coarse Data Fusion: Imaging and clinical data are often combined superficially, missing complex cross-modal interactions.
  3. Data Scarcity: Longitudinal imaging data is rarely available in large quantities, making it difficult to train models on actual disease progression.

Methodology: The DyPro Framework

The DyPro architecture addresses these challenges through a three-stage process:

  1. Patient-Specific Tumor Mapping: CT scans are segmented into five anatomical regions (parenchyma, tumors, hepatic veins, portal veins, and remnant liver). These regions, along with clinical variables, are encoded as nodes in a heterogeneous graph.
  2. Latent Residual Evolution: To overcome the lack of longitudinal data, DyPro employs a module that "rolls the graph forward" in time. Using a Graph Neural Network (GNN) as a residual operator, it simulates a 12-step sequence of future latent states, modeling the evolving relationship between the tumor and the liver's vascular structure.
  3. Cascaded Outcome Prediction: The simulated trajectory is processed by an LSTM. The model uses a cascaded design where recurrence (DFS) prediction informs the overall survival (OS) prediction, mirroring the clinical reality that recurrence often precedes death.
The DyPro framework architecture, illustrating the transition from patient feature mapping to outcome prediction.

Dataset

The model was evaluated using the MSKCC CRLM dataset, comprising 197 patients with preoperative contrast-enhanced CT scans, clinicopathologic variables (e.g., CEA levels, tumor burden), and long-term follow-up data.

Key Results and Insights

DyPro significantly outperformed traditional clinical metrics and standard AI baselines. The framework achieved a C-index of 0.755 for OS and 0.714 for DFS, vastly exceeding the performance of the Clinical Risk Score (CRS, 0.52) and Tumor Burden Score (TBS, 0.57) currently used in clinical practice.

Performance comparison of different graph backbones for DFS and OS prediction tasks.

Key insights from the study include:

  • Inferred Dynamics: The model successfully inferred plausible progression trajectories from static scans, providing stronger prognostic signals than static representations.
  • Spatial Context: Representing the liver as a graph of anatomical structures captured critical spatial relationships, such as tumor proximity to major vessels.
  • Task Conditioning: Explicitly modeling the dependency between recurrence and survival (DFS → OS) improved the accuracy of survival predictions compared to treating them as independent tasks.

Conclusion and Future Work

DyPro demonstrates the potential of dynamic graph modeling in oncology. However, the study's reliance on a single-center dataset of 197 patients suggests a need for larger, multi-center validation. Future research could integrate finer-grained data, such as genomic markers (e.g., KRAS mutations), and focus on improving the interpretability of the latent evolution states to enhance clinical adoption.

Summary updated

BibTeX
@misc{zhu2025stg,
  author = {Yiran Zhu and Wei Yang and Yan Su and Zesheng Li and Chengchang Pan and Honggang Qi},
  title = {{STG: Spatiotemporal Graph Neural Network with Fusion and Spatiotemporal Decoupling Learning for Prognostic Prediction of Colorectal Cancer Liver Metastasis}},
  year = {2025},
  eprint = {2505.03123},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.03123},
}
Figure