Spatiotemporal Graph Neural Network Modelling Perfusion MRI

TL;DR

The paper tackles non-invasive IDH mutation prediction in glioma from four-dimensional perfusion MRI (pMRI), addressing limitations of parametric perfusion models and prior spatiotemporal methods. It introduces PerfGAT, a spatiotemporal graph neural network that builds a perfusion graph from DSC-MRI using a brain atlas and tumor masks, with temporal graphs $G^T=(X^B,A^T)$ and spatial graphs $G^S=(X^B,A^S)$, and a graph structure learning module employing edge attention and a negative graph $G^{T^-}$ to refine temporal connectivity. A dual-attention feature fusion module combines temporal embeddings, spatial embeddings, and local tumor features, while a class-balanced recombining augmentation improves minority-class generalization during retraining. On a meta-cohort of 444 glioma patients, PerfGAT achieves superior accuracy, balanced accuracy, and AUC compared with state-of-the-art baselines, demonstrating effective modelling of pMRI for genotype prediction and prognosis, with future work directed at multi-modal graph analyses and clinical validation.

Abstract

Perfusion MRI (pMRI) offers valuable insights into tumor vascularity and promises to predict tumor genotypes, thus benefiting prognosis for glioma patients, yet effective models tailored to 4D pMRI are still lacking. This study presents the first attempt to model 4D pMRI using a GNN-based spatiotemporal model PerfGAT, integrating spatial information and temporal kinetics to predict Isocitrate DeHydrogenase (IDH) mutation status in glioma patients. Specifically, we propose a graph structure learning approach based on edge attention and negative graphs to optimize temporal correlations modeling. Moreover, we design a dual-attention feature fusion module to integrate spatiotemporal features while addressing tumor-related brain regions. Further, we develop a class-balanced augmentation methods tailored to spatiotemporal data, which could mitigate the common label imbalance issue in clinical datasets. Our experimental results demonstrate that the proposed method outperforms other state-of-the-art approaches, promising to model pMRI effectively for patient characterization.

Figures (2)

• Figure 1: Model Design. Top. Training Pipeline. DSC-MRI data, brain atlas and tumor masks construct input graphs. Features of focal tumor $u^I$, spatial graph $\mathbf{Z'}^{\Phi_S}$, and temporal graph $\mathbf{Z'}^{\Phi_T}$ are extracted by encoders. During retraining, features are augmented before dual-attention fusion. Resultant fused feature $\mathbf{Z}$ is used for IDH prediction. Bottom left. Temporal encoder. Pairs of $G^T$ and its negative graph $G^{T'}$ are fed into the DyGATConv module for iterative connection adjustment based on the edge attention. Bottom right. Dual-attention Feature Fusion. $u^I$ and $\mathbf{Z'}^{\Phi_S}$ are fused using node attention. The resulting weighted spatial node embedding is then fused with $\mathbf{Z'}^{\Phi_T}$ using semantic attention, resulting in fused feature $\mathbf{Z}$ by global pooling.
• Figure 2: ROC curves of comparison models (left), and ablation models (right)