4579. Impact of Deep Learning Reconstruction Algorithm on Diffusion Tensor Imaging of the Growth Plate on Different Spatial Resolutions
Authors* Denotes Presenting Author
Laura Santos;
Columbia University Irving Medical Center
Hao-Yun Hsu *;
Columbia University Irving Medical Center
Sachin Jambawalikar;
Columbia University Irving Medical Center
Jaemin Shin;
General Electric
Maggie Fung;
General Electric
Sogol Mostoufi-Moab;
Children's Hospital of Philadelphia
Diego Jaramillo;
Columbia University Irving Medical Center
Objective:
This study aims to quantitatively evaluate how deep learning (DL)-based reconstruction impacts diffusion tensor imaging metrics (DTI) on same subjects scanned with different spatial resolutions.
Materials and Methods:
Two DTI knee sequences were acquired in 22 healthy children on a 3-T GE MRI scanner using T/R knee coil, with two different voxel dimensions, and a fat-suppressed single-shot spin-echo echo-planar sequence. Parameters included: FOV, 25 cm; repetition time, 4000 ms; echo time of 50 ms, and 20 noncollinear diffusion directions; b-values: 0 and 600 sec/mm2. Voxel dimensions were: 2.0 mm 3 voxel (<em>n</em> = 22), and 2 x 2 x 3.0 mm<sup>3</sup> (<em>n</em> = 22) with no inter-section gaps on either sequence. Same raw knee DTI scans were reconstructed twice using conventional reconstruction and AIR Recon DL algorithm (GE Healthcare, Waukesha, WI) and labeled as non-DL and DL recon images. Regions of interest (ROIs) were manually drawn on DL recon images in the distal femur and proximal tibial physes using fiber tract reconstruction software TrackVis. The same ROIs were used consistently for non-DL recon images. Diffusion metrics (tract count, volume, length, and fractional anisotropy (FA) for both physes were compared in DL versus non-DL recon images. Differences were assessed using Wilcoxon-signed ranked test and Bland-Altman plots.
Results:
Femur and tibia tract count, volume, and length increased in DL recon images for both voxel dimensions, with greater increase on isovolumetric 2 mm<sup>3</sup> scans (<em>p</em> = 0.04). All diffusion metrics were significantly different between DL and non-DL recon for both physes with a 2 mm<sup>3</sup> voxel dimension (<em>p</em> < 0.001). There were no differences in femur and tibial tract count or volume in the DL versus non-DL cases using 2 x 2 x 3 mm voxels (<em>p</em> = 0.13), but tract length was different (<em>p</em> < 0.01). DL recon resulted in a significant decrease in femorotibial FA for both voxel dimensions (<em>p</em> < 0.0001). DL-based reconstruction algorithms improved accuracy of femur and tibia diffusion metrics by reducing noise inherently present in DTI MRI voxels, thereby increasing signal to noise ratios (SNRs). Smaller voxel volumes increase spatial resolution and reduce partial volume effects (PVEs) but result in lower SNR. Larger voxel dimensions increase SNRs but lower spatial resolution, increasing the likelihood of PVEs, whether DL recon or not. By leveraging DL recon algorithms on smaller voxel volumes acquisitions, noise is reduced while preserving higher spatial resolution, allowing more accurate quantification of diffusion metrics.
Conclusion:
Achieving higher spatial resolution requires smaller voxel sizes, which improve structure discrimination. DL reconstruction algorithms can then capitalize on this by reducing noise and improving SNR for DTI images.
