Deep Constrained Spherical Deconvolution for Robust Harmonization.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Data Science

Background:

• Diffusion weighted magnetic resonance imaging (DW-MRI) is crucial for studying tissue microarchitecture.
• Large-scale, multi-site DW-MRI datasets are increasingly available, but suffer from significant measurement variability.
• This variability hinders reproducible and robust analyses in multi-site and longitudinal studies.

Purpose of the Study:

• To develop a novel deep learning method for harmonizing DW-MRI signals.
• To improve the reproducibility and robustness of microstructure estimation from multi-site DW-MRI data.
• To introduce a data-driven, scanner-invariant regularization for enhanced fiber orientation distribution function (FODF) estimation.

Main Methods:

• A deep learning-based harmonization method was proposed, incorporating data-driven, scanner-invariant regularization.
• The method utilizes 8th order spherical harmonics coefficients for data representation.
• Validation was performed on the Human Connectome Project (HCP) test-retest dataset and the MASiVar dataset (inter- and intra-site).

Main Results:

• The proposed harmonization method achieved higher angular correlation coefficients (ACC) with ground truth signals (0.954 vs. 0.942) compared to a baseline supervised deep learning scheme.
• It demonstrated superior consistency of FODF signals for intra-scanner data (0.891 vs. 0.826).
• Results indicate improved accuracy and consistency in DW-MRI data harmonization.

Conclusions:

• The novel deep learning approach effectively harmonizes DW-MRI signals, enhancing microstructure estimation robustness and reproducibility.
• The data-driven framework offers a flexible solution for mitigating multi-site variability in neuroimaging.
• This method has the potential for broader application in harmonizing diverse neuroimaging datasets.