A Sparse Volume Reconstruction Method for Fetal Brain MRI Using Adaptive Kernel Regression

Area of Science:

• Medical Imaging
• Neuroimaging
• Image Reconstruction

Background:

• Slice-to-volume reconstruction (SVR) effectively reconstructs 3D fetal brain MRI but struggles with sparse sampling.
• Motion artifacts are a common challenge in fetal MRI acquisition.

Purpose of the Study:

• To develop an enhanced SVR method for sparse fetal brain MRI reconstruction from motion-corrupted data.
• To improve image quality and detailed preservation in 3D fetal brain MRI.

Main Methods:

• Incorporated slice-to-volume 2D/3D registration and point spread function (PSF)-based volume updates within the SVR framework.
• Introduced adaptive kernel regression for sparse data handling and local structure preservation using covariance matrices.
• Implemented a GPU-based acceleration for computational efficiency.

Main Results:

• Adaptive kernel regression significantly improved image quality for sparse sampling data.
• Optimal parameters identified: structure sensitivity 0.4, steering kernel size 7x7x7, and steering smoothing bandwidth 0.5.
• The GPU-based method achieved over 90x speedup compared to CPU implementation.

Conclusions:

• The proposed SVR method with adaptive kernel regression effectively reconstructs high-quality 3D fetal brain MRI from sparse, motion-corrupted data.
• This approach enhances detailed preservation and offers significant computational speed improvements.
• The method demonstrates statistical improvements in image quality for sparse sampling scenarios.