DMVFC: Deep Learning Based Functionally Consistent Tractography Fiber Clustering Using Multimodal Diffusion MRI and Functional MRI

TL;DR

DMVFC tackles the problem of fiber clustering in tractography by integrating geometry from diffusion MRI with functional signals from fMRI and microstructural FA to produce functionally coherent white matter parcellations. The method employs a two-stage training pipeline: multi-view pretraining with parallel DGCNNs for geometry and function, followed by collaborative fine-tuning to fuse modalities, and an inference stage that incorporates FA. Across 100 HCP-YA subjects and 72 WM bundles, DMVFC outperforms state-of-the-art baselines in both functional homogeneity and geometric coherence, with ablations confirming the benefits of multimodal integration and informed centroid initialization. This work demonstrates the value of WM BOLD signals for tractography clustering and points to rich future directions for incorporating additional modalities and alternative tractography pipelines to further improve reproducibility and biological interpretability.

Abstract

Tractography fiber clustering using diffusion MRI (dMRI) is a crucial method for white matter (WM) parcellation to enable analysis of brains structural connectivity in health and disease. Current fiber clustering strategies primarily use the fiber geometric characteristics (i.e., the spatial trajectories) to group similar fibers into clusters, while neglecting the functional and microstructural information of the fiber tracts. There is increasing evidence that neural activity in the WM can be measured using functional MRI (fMRI), providing potentially valuable multimodal information for fiber clustering to enhance its functional coherence. Furthermore, microstructural features such as fractional anisotropy (FA) can be computed from dMRI as additional information to ensure the anatomical coherence of the clusters. In this paper, we develop a novel deep learning fiber clustering framework, namely Deep Multi-view Fiber Clustering (DMVFC), which uses joint multi-modal dMRI and fMRI data to enable functionally consistent WM parcellation. DMVFC can effectively integrate the geometric and microstructural characteristics of the WM fibers with the fMRI BOLD signals along the fiber tracts. DMVFC includes two major components: (1) a multi-view pretraining module to compute embedding features from each source of information separately, including fiber geometry, microstructure measures, and functional signals, and (2) a collaborative fine-tuning module to simultaneously refine the differences of embeddings. In the experiments, we compare DMVFC with two state-of-the-art fiber clustering methods and demonstrate superior performance in achieving functionally meaningful and consistent WM parcellation results.

Figures (5)

• Figure 1: Method Overview. The training stage consists of pretraining and tuning processes. During the pretraining stage, two parallel feature extraction models are trained to compute embeddings from fiber geometric information and brain functional signals, respectively. During the fine-tuning stage, the pretrained embeddings are optimized to ensure that the clustering outcomes integrate both geometric and functional information simultaneously. During the inference stage, fiber geometry information and FA data are incorporated to optimize clustering results. During this phase, an input fiber is assigned to the cluster with the highest soft label assignment probability, indicating the specific cluster to which the fiber belongs.
• Figure 2: Flowchart of data preprocessing
• Figure 3: Impact of incorporating fMRI and FA information on model performance
• Figure 4: Visualization of clustering results for four different methods, with clusters colored according to the strength of fMRI signals. Each column displays the corresponding cluster from three different clustering methods. Differences between clusters are highlighted by red circles. Note that these are selected example clusters for visualization purposes from all processed bundles, not exhaustive representations.
• Figure 5: Illustration of the geometric embedding using UMAP, with the baseline DFC on the left and our proposed framework in the middle, with the visualization of tractography of the corresponding bundle. Key differences are highlighted by the red circles. The blue streamlines represent those unique fiber to our method's cluster, while red streamlines indicate those exclusive to the baseline. Yellow streamlines denote the overlap, appearing in both our cluster and the baseline cluster. From all clustered white matter tracts, we select the CC_5 (top) and SLF_III (bottom) for demonstration.