TractShapeNet: Efficient Multi-Shape Learning with 3D Tractography Point Clouds

TL;DR

This study tackles the slow computation of tractography-based shape measures by introducing TractShapeNet, a deep learning framework that regresses five shape descriptors ($length$, $span$, $volume$, $total\_surface\_area$, $irregularity$) directly from per-cluster tractography point clouds. The model employs a Siamese PointNet architecture with a hybrid loss $L_{total}=L_1+L_2+\alpha\times L_{SF}$, where $L_{SF}$ is a Fourier-domain cross-measure loss, enabling effective multi-shape learning. On a large HCP-YA dataset, TractShapeNet outperforms baseline point-cloud models in both Pearson correlation $r$ (average $r=0.895$) and normalized MSE ($nMSE=0.195$), while delivering faster inference than voxel-based tools like DSI-Studio and competitive runtimes versus other deep models. In downstream language-cognition prediction tasks, the shape measures from TractShapeNet retain predictive utility comparable to those derived from DSI-Studio, supporting the practical impact of efficient, geometry-based tractography analysis. The work thus enables scalable, accurate tractography shape analyses on large cohorts with available code.

Abstract

Brain imaging studies have demonstrated that diffusion MRI tractography geometric shape descriptors can inform the study of the brain's white matter pathways and their relationship to brain function. In this work, we investigate the possibility of utilizing a deep learning model to compute shape measures of the brain's white matter connections. We introduce a novel framework, TractShapeNet, that leverages a point cloud representation of tractography to compute five shape measures: length, span, volume, total surface area, and irregularity. We assess the performance of the method on a large dataset including 1065 healthy young adults. Experiments for shape measure computation demonstrate that our proposed TractShapeNet outperforms other point cloud-based neural network models in both the Pearson correlation coefficient and normalized error metrics. We compare the inference runtime results with the conventional shape computation tool DSI-Studio. Our results demonstrate that a deep learning approach enables faster and more efficient shape measure computation. We also conduct experiments on two downstream language cognition prediction tasks, showing that shape measures from TractShapeNet perform similarly to those computed by DSI-Studio. Our code will be available at: https://github.com/SlicerDMRI/TractShapeNet.

Figures (2)

• Figure 1: An example white matter connection, the cingulum bundle (CB, top image), extracted from the entire white matter of the human brain using an atlas-based fiber clustering approach Zhang2018-ez. The streamlines of the 22 fiber clusters in the CB are converted to point cloud representations (bottom).
• Figure 2: Overview of TractShapeNet.