[Paper Review] Tractogram filtering of anatomically non-plausible fibers with geometric deep learning

Pietro Astolfi, Ruben Verhagen+5 more|arXiv (Cornell University)|Mar 24, 2020

Advanced Neuroimaging Techniques and Applications25 references18 citations

TL;DR

This paper proposes a geometric deep learning approach using a modified Dynamic Edge Convolution (sDEC) model to filter anatomically non-plausible fibers from tractograms by treating each streamline as a point cloud and leveraging local geometric relationships. The method achieves over 95% accuracy in distinguishing plausible from non-plausible fibers, with inference on a 1M-fiber tractogram taking under 50 seconds on a GPU, outperforming baseline models in robustness to curvature and length variations.

ABSTRACT

Tractograms are virtual representations of the white matter fibers of the brain. They are of primary interest for tasks like presurgical planning, and investigation of neuroplasticity or brain disorders. Each tractogram is composed of millions of fibers encoded as 3D polylines. Unfortunately, a large portion of those fibers are not anatomically plausible and can be considered artifacts of the tracking algorithms. Common methods for tractogram filtering are based on signal reconstruction, a principled approach, but unable to consider the knowledge of brain anatomy. In this work, we address the problem of tractogram filtering as a supervised learning problem by exploiting the ground truth annotations obtained with a recent heuristic method, which labels fibers as either anatomically plausible or non-plausible according to well-established anatomical properties. The intuitive idea is to model a fiber as a point cloud and the goal is to investigate whether and how a geometric deep learning model might capture its anatomical properties. Our contribution is an extension of the Dynamic Edge Convolution model that exploits the sequential relations of points in a fiber and discriminates with high accuracy plausible/non-plausible fibers.

Motivation & Objective

• To address the persistent challenge of artifactual, anatomically non-plausible fibers in tractograms generated by diffusion MRI fiber tracking.
• To explore whether geometric deep learning models can effectively learn anatomical plausibility from point cloud representations of streamlines.
• To develop a fast, accurate, and generalizable filtering method that outperforms traditional signal reconstruction and sequential models.
• To investigate the impact of streamline length and curvature on classification performance and model robustness.

Proposed method

• The method treats each streamline as a 3D point cloud and employs a modified Dynamic Edge Convolution (sDEC) model to capture local geometric relationships between points.
• The sDEC model is designed to be permutation-invariant and order-robust, preserving performance regardless of point sequence in the streamline.
• A rule-based heuristic method (ExTractor) provides ground truth labels for training, classifying fibers as anatomically plausible (P) or non-plausible (nP).
• The model is trained in a supervised fashion using these labels, with performance evaluated via accuracy, precision, recall, and F1-score across multiple subjects.
• The approach is compared against a bidirectional LSTM (bLSTM) and a standard PointNet (PN) model to assess improvements in flexibility and accuracy.
• The method is deployed as a BrainLife app and open-sourced for reproducibility, including both code and a dataset of 20 labeled tractograms.

Experimental results

Research questions

• RQ1 Can geometric deep learning models effectively learn anatomical plausibility from 3D point cloud representations of streamlines?
• RQ2 How does the performance of sDEC compare to bLSTM and PointNet in filtering non-plausible fibers across varying streamline lengths and curvatures?
• RQ3 To what extent does the model’s robustness to point ordering and geometric structure improve classification accuracy?
• RQ4 How do false positive and false negative error distributions vary across different streamline morphologies?

Key findings

• The sDEC model achieves a mean accuracy of over 95% in distinguishing anatomically plausible from non-plausible fibers, with low standard deviation across subjects.
• The model filters a whole tractogram of approximately 1 million streamlines in under 46.2 seconds on an NVIDIA Titan Xp GPU.
• sDEC outperforms both bLSTM and PointNet, particularly in handling long and highly curved streamlines, where it reduces false positive errors by 40% compared to PN.
• The error analysis reveals that PN exhibits a significant bias toward classifying long and curved streamlines as plausible, leading to higher false positive rates.
• The model demonstrates strong generalization, remaining accurate even when faced with noisy or inconsistent ground truth labels from the ExTractor heuristic method.
• The sDEC model’s ability to model local geometric edges between points provides a competitive advantage in complex, curved pathways.