Introduction: Diffusion Tensor Imaging (DTI) quantitative metrics, such as Fractional Anisotropy (FA) and Mean Diffusivity (MD), offer objective measurements that are crucial for detecting and understanding pathological changes in the brain’s microstructure. However, the accuracy of DTI metrics is compromised by systematic errors, including those caused by nonlinearity of the gradient magnetic field, leading to spatial distortions (1). These influences the objectivity and comparability of DTI measurements, potentially impacting clinical and research outcomes.

To address these systematic errors, the BSD-DTI method could be employed (2). This involves acquiring DTI data from a phantom using the same protocol as the target measurement. This approach allows for the precise calculation of a spatially heterogeneous b-matrix, which can then be used to accurately compute the DTI for in vivo measurements (3). Despite its effectiveness, this method is time-consuming, as it requires phantom DTI acquisition each time. To streamline the process, efforts have been made to develop a solution that computes spatial b-matrices without the need for an additional phantom measurement.

Methods: In our work, we present a results of deep learning model trained to generate a spatially varying b-matrix, with reference data obtained using the BSD-DTI method. The model (AIBSD) utilized a basic convolutional neural network. Its input comprised: a non-diffusion-weighted (b = 0) image, a diffusion-weighted image acquired along a single direction, the components of the gradient direction vector, and the value of the b-factor. The model’s output consisted of three b-matrices, each representing one of the three diffusion directions. The dataset contained 130 studies, each including both phantom and in vivo measurements. The model was provided with input sourced from either phantom measurements or in vivo measurements.

The results were evaluated using an isotropic phantom. For such a phantom, an ideal measurement would yield an FA value of 0. The b-matrices were generated from both phantom (AIBSD_P) and in vivo (AIBSD_B) measurements. Results were compared with standard DTI without systematic error removal (STD) and with the BSD-DTI method (BSD).

Results: The obtained results indicate that the FA values on phantoms are significantly lower using the BSD and AIBSD methods compared to the STD method, with values closer to the ideal of 0. Specifically, FA values are 4.03E-02 (STD), 2.88E-02 (BSD), 2.90E-02 (AIBSD_P), and 3.06E-02 (AIBSD_B). The AIBSD_B method is slightly less efficient in removing bias compared to BSD and AIBSD_P, but still removes it to a significant extent, reducing FA by approximately 24% compared to STD. Meanwhile, MD values remain consistent across all methods: 2.07E-03 (STD), 2.06E-03 (BSD), 2.05E-03 (AIBSD_P), and 2.05E-03 (AIBSD_B).  

Conclusions:

The analysis shows that AIBSD methods significantly reduce systematic errors in FA measurements, matching the performance of the BSD method.  MD values remain consistent, demonstrating robustness. B matrices generated from the brain exhibit high potential for further minimizing systematic errors and could potentially allow systematic errors to be removed without the use of phantoms. However, additional measurements on in vivo studies are required to confirm these findings. Overall, AIBSD methods enhance FA accuracy without compromising MD measurements, proving valuable for improved DTI metrics accuracy.

Acknowledgments

Funding: Medical Research Agency contract no:2020/ABM/01/00006-00

Image/Figure:
Click to View

Image/Figure Caption:

The figure presents a comparison of FA metrics obtained from an isotropic phantom using four methods (STD, BSD, AIBSD_P, and AIBSD_B). Diagonal elements display histograms, the lower left quadrant features Bland-Altman plots, and the upper right quadrant shows data fitted with a linear regression model.

Author