Diffusion MRI based tractography has become the primary tool to construct structural brain network for human connectome analysis. To achieve plausible results, it is critical to optimize the imaging protocol for an appropriate algorithm. Many fiber tracking algorithms have been developed over the last two decades. Due to the complexity of brain microstructure and organization, the validation of tractography results remains a challenging problem. In this work, the effect of signal-to-noise ratio (SNR) and b value on fiber tracking were evaluated for three algorithms through a software phantom developed by Tractometer (tractometer.org/).

A phantom of 26 white matter bundles provided by Tractometer was used as ground truth. The diffusion MRI signal was generated using Fiberfox¹ tool implemented in the diffusion module of MITK (http://mitk.org/) and then exported to different tractography software. By using Fiberfox, raw diffusion MRI datasets with different noise and b-values ware simulated from the phantom. Two b values were set for the simulation: b = 1000 s/mm2, b = 2000 s/mm2. For each acquisition, three different noise levels were assigned, resulting a final signal-to-noise ratio (SNR) of 22, 32, and 44. Each diffusion dataset contains 32 gradient directions and three b0 images. The image matrix for all datasets is 90 x 108 x 90 and an isotropic resolution of 2 mm and TR/TE = 4000/108 ms. Motion effect, ghost, and susceptibility distortion artifacts were also added in the simulation.

The fiber tracking was performed with three tractography algorithms: Diffusion tensor imaging (DTI), constrained spherical deconvolution (CSD)² and generalized q-sampling imaging (GQI)³ using Diffusion Toolkit (trackvis.org), MRtrix (https://www.nitrc.org/projects/mrtrix/) and DSI Studio (http://dsi-studio.labsolver.org) respectively. Fractional anisotropy (FA) was first computed for b = 1000 and SNR = 44 to generate a seeding mask with FA > 0.2. The tracking parameters of DTI as well as GQI included an angle threshold = 35˚. The FACT⁴ algorithm was applied for DTI. For CSD, the deterministic streamline technique was employed at order 6. For each algorithm and diffusion dataset, the fiber tracking was performed twice to obtain two tractography files. Seed density was adjusted for DTI to generate roughly the same number of streamlines as the ground truth data (~200000).

The cortical region of the brain was parcellated into 278 ROIs⁵. Number of fibers between any pair of ROIs were computed to form a connectome matrix. Two analysis were performed. First, the variability of tracking algorithm was evaluated by computing the standard deviation of the difference between the two connectome matrices6. Second, the connectome was binarized and compared against the ground truth binarized connectome. The binarization entailed thresholding the connectome by removing connections if the number of fibers is smaller than a certain number. The false positive rate (FPR, ratio of spurious connections to the total number of true connections) and the false negative rate (FNR, ratio of missing connections to the total number of true connections) were computed respectively. The effect of thresholding was evaluated for FPR, FNR, and total error rate (FPR+FNR).

The distribution of fiber length for different tracking algorithms for b = 1000 and SNR = 44 is displayed in Fig. 1 along with ground truth. The results from DTI and GQI are very similar, with an overestimation of shorter fibers and underestimation of long fibers. CSD did a better job in finding longer fibers, but the whole distribution was skewed to the long fiber end. The intrinsic variability of fiber tracking is shown in Table 1. The numbers are quite close to each, without apparent effects of SNR or b values.

The false positive and false negative rate is plotted against thresholding for each condition (Fig. 2). The FPR is insensitive to SNR or b value. But in general higher SNR leads to lower FNR. In comparing the algorithms, CSD has lower FNR but higher FPR than DTI and GQI. The total error rate decreases rapidly as threshold increases but become flat at two to three times the variability of fiber tracking (see Table 1).

Large false positive and false negative rate was seen for all methods. A comparison of CSD and the other two methods suggests that the false negative is mainly caused by failure of tracking longer fibers. A moderate thresholding can reduce false positive rate and total error rate. Considering there exist other validation tools to remove false fibers such as Life⁶, it might be more desirable for a tracking algorithm to get low FNR at the first place. We will further investigate other factors such as resolution and number of diffusion gradients.