[Paper Review] Tract orientation mapping for bundle-specific tractography

Jakob Wasserthal, Peter Neher+1 more|arXiv (Cornell University)|Jun 14, 2018

Advanced Neuroimaging Techniques and Applications8 references7 citations

TL;DR

This paper introduces Tract Orientation Mapping (TOM), a deep learning approach that learns bundle-specific tractography priors by regressing from fODF peaks to orientation maps representing major white matter tracts. Using a 2D encoder-decoder FCNN, TOM achieves high-accuracy, fast (under 2 minutes per subject), and anatomically coherent tractography without whole-brain tractography or atlas registration, outperforming state-of-the-art methods in Dice score, angular error, and runtime.

ABSTRACT

While the major white matter tracts are of great interest to numerous studies in neuroscience and medicine, their manual dissection in larger cohorts from diffusion MRI tractograms is time-consuming, requires expert knowledge and is hard to reproduce. Tract orientation mapping (TOM) is a novel concept that facilitates bundle-specific tractography based on a learned mapping from the original fiber orientation distribution function (fODF) peaks to a list of tract orientation maps (also abbr. TOM). Each TOM represents one of the known tracts with each voxel containing no more than one orientation vector. TOMs can act as a prior or even as direct input for tractography. We use an encoder-decoder fully-convolutional neural network architecture to learn the required mapping. In comparison to previous concepts for the reconstruction of specific bundles, the presented one avoids various cumbersome processing steps like whole brain tractography, atlas registration or clustering. We compare it to four state of the art bundle recognition methods on 20 different bundles in a total of 105 subjects from the Human Connectome Project. Results are anatomically convincing even for difficult tracts, while reaching low angular errors, unprecedented runtimes and top accuracy values (Dice). Our code and our data are openly available.

Motivation & Objective

• To address the challenge of manual, time-consuming, and non-reproducible dissection of white matter tracts in diffusion MRI.
• To eliminate reliance on whole-brain tractography, atlas registration, and clustering by learning direct tract-specific orientation priors.
• To develop a deep learning framework that maps fODF peaks to tract-specific orientation maps (TOMs) for use as priors or direct inputs in tractography.
• To establish a new benchmark in accuracy, runtime, and anatomical plausibility for bundle-specific tractography.

Proposed method

• A fully convolutional neural network (FCNN) with encoder-decoder architecture and skip connections is trained to map 3D fODF peak images (9 input channels) to 20 tract-specific orientation maps (60 output channels).
• The network uses 2D slices along the y-axis to reduce memory usage while preserving spatial resolution, with input fODF peaks derived from constrained spherical deconvolution (CSD).
• A custom loss function combines weighted cosine similarity and mean squared error of l2-norms to handle class imbalance between background and tract voxels.
• The model is trained on 105 HCP subjects with semi-automatically generated reference tracts, using a learning rate of 0.001 and Adamax optimization.
• TOM outputs are used either as a prior for tractography or directly as orientation fields, enabling bundle-specific streamline tracking.
• The method avoids atlas registration, clustering, and whole-brain tractography, streamlining the pipeline.

Experimental results

Research questions

• RQ1 Can a deep learning model learn accurate, tract-specific orientation maps directly from fODF peaks without whole-brain tractography?
• RQ2 How does TOM compare to state-of-the-art methods in terms of Dice score, angular error, and runtime for 20 major white matter tracts?
• RQ3 Can TOM produce anatomically coherent and complete bundles even for complex, overlapping tracts?
• RQ4 Does using TOM as a prior or direct input improve tractography accuracy compared to using fODF peaks alone?

Key findings

• TOM achieved the highest Dice scores across all 20 bundles, outperforming RecoBundles, TractQuerier, WMA, and Atlas Registration in quantitative evaluation.
• The method achieved low angular errors, with predicted TOM peaks showing lower angular deviation from reference peaks than the original fODF peaks.
• TOM completed bundle reconstruction in under 2 minutes per subject, making it more than five times faster than the second-fastest method (Atlas Registration, ~11 minutes).
• TOM-based tractography produced spatially coherent, complete bundles without false positives, while RecoBundles on deterministic tractograms missed critical parts like lateral CST projections.
• Direct TOM tractography yielded smoother and more accurate results than peak selection from TOM, aligning with lower angular errors.
• The method demonstrated robustness on challenging tracts, including the arcuate fasciculus and corticospinal tracts, with minimal false positives.