Registration and quantification network (RQnet) for IVIM-DKI analysis in MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Diffusion MRI analysis is challenged by registration errors due to varying DWI contrasts.
• Accurate registration and quantification are crucial for intravoxel incoherent motion-diffusion kurtosis imaging (IVIM-DKI).

Purpose of the Study:

• To develop an unsupervised deep learning method for simultaneous DWI registration and IVIM-DKI parameter quantification.
• To address registration inaccuracies in diffusion MRI by proposing a novel deep learning approach.

Main Methods:

• An unsupervised deep learning framework was designed for registration and quantification of IVIM-DKI parameters.
• Motion-simulated data from 17 healthy volunteers and 4 subjects with head motion were used for training and testing.
• Kidney images were acquired to assess applicability to other organs.
• Compared registration accuracy against Statistical Parametric Mapping and a normalized cross-correlation loss deep learning method.
• Quantification utilized a deep learning method incorporating diffusion gradient direction information.

Main Results:

• The proposed method demonstrated accurate registration and quantification for IVIM-DKI analysis in simulations and experiments.
• High registration accuracy was achieved across all b-values.
• The method outperformed compared techniques in quantification performance during in vivo experiments.

Conclusions:

• The developed deep learning method effectively aligns DWIs and accurately quantifies IVIM-DKI parameters.
• This approach offers improved accuracy for diffusion MRI analysis, particularly for IVIM-DKI studies.