AIBSD: Deep learning approach to address spatial systematic errors in diffusion tensor imaging

Diffusion Tensor Imaging (DTI) metrics, such as fractional anisotropy (FA) and mean diffusivity (MD), are critical for assessing tissue microstructure but are susceptible to spatial systematic errors caused by inaccuracies in b-matrix estimation. Traditional correction methods require time-consuming phantom measurements, limiting their clinical utility. This study aims to develop and validate a deep learning-based method, AIBSD, to generate spatially corrected b-matrices for DTI, eliminating systematic errors without additional phantom scans.

A retrospective analysis of 130 DTI datasets, including paired in vivo and phantom measurements, was conducted. A convolutional neural network (CNN) with a dual-encoder architecture was trained to predict corrected b-matrices using reference data generated by the BSD-DTI method. DTI metrics were computed using four approaches: standard (STD), BSD-DTI (BSD), AIBSD with phantom data (AIBSD_P), and AIBSD with in vivo data (AIBSD_B). Agreement between methods was quantified using Lin's concordance correlation coefficient (CCC).

For the isotropic phantom, FA values decreased significantly with corrected b-matrices: 0.0407 (STD) vs. 0.0291 (BSD), 0.0289 (AIBSD_P), and 0.0311 (AIBSD_B), demonstrating the ability of these methods to reduce systematic errors. In vivo analysis of deep gray matter demonstrated strong agreement between BSD and AIBSD_P (CCC = 0.998) and AIBSD_B (CCC = 0.993), compared to lower agreement with STD (CCC = 0.861). AIBSD_B achieved comparable accuracy to BSD without requiring phantom data.

AIBSD effectively eliminates spatial systematic errors in DTI by generating corrected b-matrices directly from in vivo data, bypassing the need for additional phantom scans. The method demonstrates high concordance with BSD-DTI and improves the accuracy of FA and MD metrics. This advancement enhances the feasibility of precise DTI in clinical and research settings, streamlining workflows while maintaining reliability. This study is the first attempt to use deep neural networks to address the fundamental problem of estimating and correcting the influence on the measurement of the true spatial distribution of magnetic field gradients present during DTI.