Improved diffusion tensor parametric maps in the brain from reduced-direction acquisitions using a deep learning approach.

OBJECTIVES: This study implements and evaluates a deep learning approach to enhance FA, MD, AD, and RD maps derived from 16-direction DTI to approximate corresponding maps derived from 32-direction DTI, with the aim of reducing acquisition time while maintaining quantitative accuracy. MATERIALS AND METHODS: A modified 8-layer 3D convolutional neural network was trained using scalar DTI-derived maps from 16-direction acquisitions as input and corresponding 32-direction maps as reference targets. Model performance was evaluated on an independently acquired evaluation cohort (n=33) using PSNR, %NRMSE, SSIM, Bland-Altman analysis, and tract-based spatial statistics (TBSS), with BM4D and AONLM as comparators. RESULTS: The DL model reduced noise and improved PSNR, %NRMSE, and SSIM compared with raw ND16, BM4D, and AONLM (p < 0.05). DL-derived FA, MD, and AD showed the smallest deviations from the ND32 reference among the evaluated methods, although statistically significant mean differences remained. TBSS showed localized differences for FA and AD, but no significant differences for MD or RD (p > 0.05). DISCUSSION: The proposed map-domain DL approach improved scalar DTI-derived maps from 16-direction acquisitions while reducing diffusion acquisition time by approximately 1.8-fold. The findings demonstrate improved agreement with 32-direction reference maps, while remaining quantitative differences and generalizability require further validation.