Khái niệm cốt lõi
GDCNet demonstrates fast distortion correction of functional images using deep learning, outperforming traditional methods.
Tóm tắt
GDCNet is a novel approach for geometric distortion correction in fMRI images without the need for additional sequences. It leverages T1-weighted anatomical images for distortion correction, reducing scan time and improving efficiency. The self-supervised models achieved the best distortion correction performance across different datasets. These models showed significant improvements in normalized mutual information (NMI) compared to benchmark methods like FUGUE and TOPUP. The GDCNet models also exhibited faster processing speeds, eliminating the need for additional sequence acquisitions for VDM estimation. The study evaluated the models on retrospective and prospective datasets, demonstrating successful distortion correction even with low PE bandwidths.
Thống kê
GDCNet demonstrated processing speeds 14 times faster than TOPUP in the prospective dataset.
The self-supervised models achieved statistically significant improvements in NMI compared to traditional methods.
Training the 3D models took less time due to processing a full time-point simultaneously.