Deep fiber clustering: Anatomically informed fiber clustering with self-supervised deep learning for fast and effective tractography parcellation

Abstract

White matter fiber clustering is an important strategy for white matter parcellation, which enables quantitative analysis of brain connections in health and disease. In combination with expert neuroanatomical labeling, data-driven white matter fiber clustering is a powerful tool for creating atlases that can model white matter anatomy across individuals. While widely used fiber clustering approaches have shown good performance using classical unsupervised machine learning techniques, recent advances in deep learning reveal a promising direction toward fast and effective fiber clustering. In this work, we propose a novel deep learning framework for white matter fiber clustering, Deep Fiber Clustering (DFC), which solves the unsupervised clustering problem as a self-supervised learning task with a domain-specific pretext task to predict pairwise fiber distances. This process learns a high-dimensional embedding feature representation for each fiber, regardless of the order of fiber points reconstructed during tractography. We design a novel network architecture that represents input fibers as point clouds and allows the incorporation of additional sources of input information from gray matter parcellation. Thus, DFC makes use of combined information about white matter fiber geometry and gray matter anatomy to improve the anatomical coherence of fiber clusters. In addition, DFC conducts outlier removal naturally by rejecting fibers with low cluster assignment probability. We evaluate DFC on three independently acquired cohorts, including data from 220 individuals across genders, ages (young and elderly adults), and different health conditions (healthy control and multiple neuropsychiatric disorders). We compare DFC to several state-of-the-art white matter fiber clustering algorithms. Experimental results demonstrate superior performance of DFC in terms of cluster compactness, generalization ability, anatomical coherence, and computational efficiency.

Keywords: Deep learning,; Fiber clustering; Image diffusion MRI; Self-supervised learning; Tractography.

Fig. 1.

Overview of our DFC framework. A self-supervised learning strategy is adopted with the pretext task of pairwise fiber distance prediction. In the pretraining stage, input fibers are encoded as embeddings with the Siamese Networks. K-means clustering is then performed on the obtained embeddings to generate initial cluster centroids. In the clustering stage, based on the neural network of the pretraining stage, a clustering layer is connected to the embedding layer and generates soft label assignment probabilities q (as shown in the orange dashed box). During training, a prediction loss (L_p) and a KL divergence loss (L_c) are combined for network optimization. During inference, an input fiber is assigned to cluster c with the maximum soft label assignment probability, which is calculated from the trained neural network. (n_p: number of points; n_e: dimension of embeddings; n_c: number of clusters).

Fig. 2.

Illustration of the process of graph construction for the input of the DGCNN model.

Fig. 3.

Visualization of example clusters from four methods (DFC, DFC_conf, WMA, QB) across three datasets (HCP, CNP, PPMI). Three example clusters are selected within known CC4, Sup-FP and AF tracts respectively. For clusters within CC4 tracts, an anterior view is displayed; for AF clusters, an inferior view is displayed; for Sup-FP clusters, a left view is displayed. (Abbreviations: CC4 - corpus callosum 4; Sup-FP - superficial-frontal-parietal; AF - arcuate fasciculus; DFC - Deep Fiber Clustering; DFC_conf - conference version of Deep Fiber Clustering; WMA - WhiteMatterAnalysis; QB - QuickBundles).

Fig. 4.

Example clusters for visualization of coherence between clusters and cortical parcels, across different methods (DFC, DFC_conf, WMA and QB) from HCP data. Clusters within the CST, IoFF and Sup_PT tracts are shown in (a), (b) and (c) respectively. For (a) and (c), the first row displays a posterior view; In the second row, the display view is indicated by the human figure at the bottom right corner; The third row is a zoomed-in area of the orange rectangle area in the second row. In (b), the first and second rows show the inferior and posterior view of the IoFF cluster. (Abbreviations: CST - corticospinal tract; IoFF - inferior occipito-frontal fasciculus; Sup_PT - superficial parieto-temporal; DFC - Deep Fiber Clustering; DFC_conf - conference version of Deep Fiber Clustering; WMA - WhiteMatterAnalysis; QB - QuickBundles).

Fig. 5.

Example tracts for visualization of cluster subdivisions within a tract, across DFC and WMA methods from HCP data. Part of the AF, CC6 and Sup_FP tracts with one or two example clusters are shown in (a), (b) and (c) respectively. The display views are indicated by the human figure at the bottom right corner. (Abbreviations: AF - arcuate fasciculus; CC6 - corpus callosum 6; Sup_FP - superficial fronto-parietal; DFC - Deep Fiber Clustering; WMA - WhiteMatterAnalysis).

Fig. 6.

Visualization of example corresponding clusters from DFC and DCEC. Colors represent the sequence number of points along the fiber (rainbow coloring with red for starting and purple for ending points).

Fig. 7.

Example clusters to compare previous (RO_absolute) and current outlier removal methods (RO_adaptive). Results of two clusters (two rows) from no outlier removal (w/o RO), RO_absolute and RO_adaptive methods are displayed in columns 1–3 respectively. The fiber color indicates the soft label assignment probability of the fiber (rainbow coloring with red indicating the smallest and purple the largest). The fourth column shows the soft label assignment probability distribution within the selected clusters. The red and green dashed lines indicate the thresholds of outlier removal for RO_adaptive and RO_absolute, respectively.