Reliability of Fiber Quantification for DTI and NODDI Metrics: A Comparison of Segmentation Methods

2174 

Abstract Submission 

Lingyu Li^1,2, Qiqi Tong³, Chenxi Lu^1,4, Hongjian He^1,5,6

¹Center for Brain Imaging Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China, ²Polytechnic Institute, Zhejiang University, Hangzhou, Zhejiang, China, ³Research Center for Healthcare Data Science, Zhejiang Lab, Hangzhou, Zhejiang, China, ⁴School of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, Zhejiang, China, ⁵School of Physics, Zhejiang University, Hangzhou, Zhejiang, China, ⁶State Key Laboratory of Brain-Machine Intelligence, Zhejiang University, Hangzhou, Zhejiang, China

Lingyu Li  
Center for Brain Imaging Science and Technology, Zhejiang University|Polytechnic Institute, Zhejiang University
Hangzhou, Zhejiang, China|Hangzhou, Zhejiang, China

Qiqi Tong  
Research Center for Healthcare Data Science, Zhejiang Lab
Hangzhou, Zhejiang, China

Chenxi Lu  
Center for Brain Imaging Science and Technology, Zhejiang University|School of Biomedical Engineering and Instrument Science, Zhejiang University
Hangzhou, Zhejiang, China|Hangzhou, Zhejiang, China

Hongjian He  
Center for Brain Imaging Science and Technology, Zhejiang University|School of Physics, Zhejiang University|State Key Laboratory of Brain-Machine Intelligence, Zhejiang University
Hangzhou, Zhejiang, China|Hangzhou, Zhejiang, China|Hangzhou, Zhejiang, China

In recent years, fiber quantification has gained popularity as a method for accurately characterizing microstructural information along fiber tracts in the individual brain (Colby et al., 2012; Yeatman et al., 2012). The fiber quantification pipeline typically involves three steps: fiber segmentation, metric estimation, and quantification. Among these steps, fiber segmentation is crucial for obtaining reliable results. In this study, we compared the reliability of quantification for microstructural metrics derived from DTI and NODDI using three established fiber segmentation methods: AFQ (Kruper et al., 2021), TractSeg (Wasserthal et al., 2018), and RecoBundle (Garyfallidis et al., 2018). Additionally, we investigated the effects of within-subject and between-subject variance. These analyses demonstrate the specificity of different segmentation strategies across different fibers and measurement metrics, highlighting the complexity of existing fiber quantification methods.

We used the HCP test-retest dataset (Van Essen et al., 2013) with a subset of 15 participants. T1-weighted MRI data and multi-shell (b = 1000, 2000, 3000 s/mm²) diffusion MRI data with 270 directions were acquired. All participants underwent test-retest measurements in the LR and RL phase-encoding directions for reliability research.
The Multi-shell multi-tissue constrained spherical deconvolution (MSMT-CSD) model was utilized to reconstruct fiber orientation distributions (FODs) using MRtrix3. Fiber segmentation was performed using three strategies: TractSeg, AFQ, and RecoBundle. Ten tracts in common were finally reconstructed, including the bilateral arcuate tracts (AF), the bilateral corticospinal tracts (CST), the bilateral inferior fronto-occipital fasciculus (IFO), the bilateral inferior longitudinal fasciculus (ILF), and the bilateral uncinate fasciculus (UF). Fractional anisotropy (FA) and neurite density index (NDI) were estimated using DIPY 1.7.0 and dmri-amico. Profiles with 100 nodes were generated between pairs of waypoints for each metric and tract by applying the AFQ algorithm using pyAFQ.
The two-way mixed effects model intraclass correlation coefficient (ICC) was calculated to assess the reliability of each position of profiles. The between-subject variance (Vb) and within-subject variance (Vw) were calculated.

As shown in Figure 1, the overall reliability of quantification was higher for FA compared to NDI. For NDI, the ICC was consistently below 0.6, indicating low reliability for most fiber quantification results. The analysis of Vb and Vw also revealed that, for FA, the inter-subject variability (Vb) was higher than the variability between different trials (Vw), whereas for NODDI, the advantage of Vb over Vw was not significant.
In terms of the FA, TractSeg demonstrates superior reliability on fiber bundles such as CST, AF, and ILF, while RecoBundle performs similarly to TractSeg on CST and AF. AFQ is more suitable for IFO and UF. Regarding the NDI, TractSeg shows superior reliability on fiber bundles like CST, AF, IFO, and ILF, while RecoBundle is more suitable for UF. The optimal fiber segmentation scheme for different tracts and different metrics was summarized in Table 1.

This study compared the reliability of fiber quantification in fiber segmentation and further analyzed the between-subject and within-subject variance in a more detailed manner. Ultimately, we obtained strategies for selecting tract-specific and metric-specific fiber segmentation methods. However, it should be noted that the reliability of NDI was generally lower than that of FA, which could be attributed to the difference in model complexity. The limited sample size of the participants may also be an important underlying factor. Overall, our findings provide evidence for the specificity of different segmentation strategies across various fibers and measurement metrics, thereby emphasizing the intricate nature of current fiber quantification methods.

Diffusion MRI Modeling and Analysis ²

White Matter Anatomy, Fiber Pathways and Connectivity ¹

Workflows

Data analysis

MRI

Tractography

White Matter

Workflows