Fig 1.
Schematic drawing of the architecture of the 1D CNN model for denoising.
“16x1x16” means the dimension of convolutional operation: kernel size = 16, depth = 1, number of filters = 16.
Fig 2.
A. We simulated a noise-free ground truth dataset (1) and added noise to create a low-noise dataset with SNR 30 (2) and a high-noise dataset with SNR 10 (3) to evaluate 1D-CNN denoising. The low-noise and high-noise simulated datasets were both denoised using 1D-CNN and MP-PCA, resulting in four additional datasets: low-noise after 1D-CNN denoising (4), low-noise after MP-PCA denoising (5), high-noise after 1D-CNN denoising (6), and high-noise after MP-PCA denoising (7). B. Denoising SoS-related noise using SENSE1 reference data. MRI scans were acquired from 3 subjects for denoising validation. There are two identical runs, each consisting of one T1w scan, two DWI scans with opposite phase encoding directions (DWI-AP and DWI-PA). Each DWI scan is constructed with SoS and SENSE1 coil combines, respectively to generate two sets of images.
Table 1.
Data, description of analyses, and web-links to the open source code and open cloud services used in the creation of this dataset can be viewed in their entirety here: https://doi.org/10.25663/brainlife.pub.35.
Fig 3.
Difference (arbitrary unit) between denoised image and original noisy image (top) and the ground truth (bottom) for MP-PCA and 1D-CNN from simulation.
Fig 4.
The noisy DWI images (SNR 10 and 30), corresponding denoised images with MP-PCA and 1D-CNN, and the ground truth, along with derived FA, MD, and ICVF maps of a representative slice from simulation.
Fig 5.
(a) A representative slice of the SoS image, the SENSE1 image, and the denoised SoS (dSoS) image for b = 1000 s/mm2 and b = 2500 s/mm2; (b) The average correlation coefficients between SENSE1 and dSoS images, and between SENSE1 and SoS images, for gray matter (GM) and white matter (WM). The dashed lines indicate the magnitude of correlation coefficients between two repeated scans with SENSE1 (SENSE1 inter-scan) as a reference.
Fig 6.
Different diffusion metrics (FA, MD, ICVF, and OD) computed from SoS, SENSE1, and dSoS images showing on a representative slice.
The corresponding histograms are shown in the bottom. For ICVF and OD, the dSoS data (green) is more similar to the gold standard SENSE1 data (blue) than the original SoS data (red). The benefit of our denoising procedure is most evident in ICVF and OD, the two measures that are derived from higher-order diffusion models that rely on high b-value images, although some improvement can also be seen in the MD value.
Table 2.
Correlations of the tensor-based diffusion metrics (i.e., FA, MD) and NODDI-based diffusion metrics (i.e., OD, ICVF) between SENSE1-SoS, SENSE1-dSoS, and SENSE1-SENSE1 repeat.
Fig 7.
The difference in the number of tractography streamlines between SENSE1 and SoS, SENSE1 and dSoS, and SENSE1 repetitions for nine bundles: IFOF (left and right), ILF (left and right), Forceps Major, Arc (left and right), and SLF 1 and 2 (left and right).
The x-axis is the average difference in streamline count between the comparison (i.e., SENSE1-SoS (red), SENSE1-dSoS (blue), SENSE1-SENSE1 repeat (black)). The error bars represent 90% confidence intervals.
Table 3.
SENSE1-dSoS correlations of diffusion metrics (FA, MD, OD, and ICVF).
For the dSoS data used in this table, the model was trained on data that was merged from the other two subjects. The dSoS data for SUBJ 1, for example, was obtained by applying a model trained on the combined data of SUBJ 2 and SUBJ 3 to SUBJ 1’s SoS data.
Fig 8.
The original high-noise image, low-noise image, and denoised images using MP-PCA and 1D-CNN along with derived MD and ICVF maps of a representative slice in Experiment 1 (iPAT2 vs.
iPAT1) for b-value 2500 s/mm2.
Fig 9.
The original high-noise image, low-noise image, and denoised images using MP-PCA and 1D-CNN along with derived MD and ICVF maps of a representative slice in Experiment 2 (SMS4 vs.
SMS1) for b-value 2500 s/mm2.