Erin J Kelly1, Hung P Do2, Dawn M Berkeley2, and Jonathan K Furuyama2
1Canon Medical Systems USA, Inc, Tustin, CA, United States, 2Canon Medical Systems USA, Inc., Tustin, CA, United States
Synopsis
dDLR
algorithms with two path CNN architecture that are designed to be noise adaptive
are robust against differences in image contrast and field
strength. This study shows clinical
image quality improvement on 1.5T brain and knee datasets using an algorithm
trained on 3T data.
Introduction
Denoising
approaches with deep learning based reconstruction (dDLR) have been shown to be
adaptive to various noise levels and contrasts of MR images (1,2). Following training and validation at 3T with
pairs of high SNR ground truth and noisy images (Ground truth + artificial
gaussian noise, which SD is [0 20](%) of max. SI), dDLR has been reported to outperform
block matching and 3D filtering (BM3D)(2). In phantom and volunteer testing, dDLR
significantly reduced image noise while preserving image quality for brain MR
images on 3T(3). dDLR uses a
two path DCNN architecture: 1- feature extraction and multiple feature
conversion layers and 2 – a separate pathway for the zero-frequency component
of the discrete cosine transformation (DCT).
Separation of this zero-frequency component from the feature extraction
layer maintains the image contrast.
Furthermore, a single neural network is trained by the entire training
set containing a variety of contrasts , therefore even the contrast difference
in the sample images have no impact on the result, since noise is being removed
regardless of contrast weighting. Given
the various types contrast possibilities of images and levels of noise in MR,
as well as changes in tissue characteristics at 1.5T compared to 3T, the
purpose of this study was to verify that dDLR can be applied to 1.5T based on
its architectural design. The goal of
this study was to determine whether the dDLR algorithm trained on 3T data can
improve image quality on clinical 1.5T brain and knee images on a qualitative
scale. Methods
Twenty
clinical brain raw data studies and twenty clinical knee raw data studies were
collected with permission from 1.5T MRI scanners (Vantage Orian, Canon Medical Systems, Tochigi, Japan), anonymized,
reconstructed using dDLR and 3 standard clinical denoising and gain algorithms
(NL2, GA43 and GA53), and randomized.
The anonymized and randomized brain and knee clinical datasets were
uploaded to an online Dicom viewer to be reviewed blindly by 3 Neuro
radiologists and 3 MSK radiologists, respectively. Each of the 20 brain studies included the
following sequences: Diffusion, T1, T2, T2*, MP-Rage, FLAIR, TOF, and post
Gd. Each of the 20 knee studies included
four copies of the following sequences: PD, PD FS, T1, T2 STIR, T2, and T2
FS. Each radiologist viewed and
qualitatively evaluated each series based on the appearance of image sharpness,
image contrast, SNR, image noise, image noise texture homogeneity,
lesion/pathology conspicuity and overall image quality according to a 1-5 Likert
scale (4). The Friedman test
was used to determine overall statistical significant difference within the
group, followed by separate post-hoc Wilcoxon signed-rank tests on the
different combinations of groups with a Bonferroni adjustment for a
significance level of 0.017 (0.05/3).
SPSS software (IBM) was used for statistical analysis. Results
Representative brain and knee images used in the
evaluation are shown in Figures 1 and 2. For the 1.5T brain evaluation, dDLR was preferred by the radiologists compared to
NL2, GA43, and GA53 according to the higher average score of DLR in all review
categories. Significant differences were found for all comparisons (p<0.0001). dDLR scored higher than the other filters in
overall image quality, Image sharpness, SNR, contrast, noise, noise texture
homogeneity, and lesion conspicuity. Average
brain image quality scores for each filter and reviewer is shown in Figure 3. For the knee evaluation, dDLR was preferred by the radiologists compared to
NL2 in all review categories according to the higher average score of dDLR
(p<0.0001). Between DLR:GA53 and
DLR:GA43, strong preferences for DLR was observed in overall image quality,
sharpness, SNR, noise, noise texture homogeneity, and lesion conspicuity
(p<0.0001). Average brain and knee overall
image quality scores for each filter and reviewer is shown in Figures 3 and 4. All
images were diagnostic. Discussion
The
dDLR training data set comprised of pairs of a high-SNR target and a noisy
input images, optimized through an offline training process, has learned to
differentiate signal from noise and can then be applied to denoising input
images regardless of contrast and field strength. dDLR can handle the differences between various
contrast weightings at 3T (3), where the contrast differences are large.
This study shows that dDLR can be applied to 1.5T with improved image quality. Any subtle contrast differences in 1.5T
images are not affected by the dDLR algorithm.
The results of this clinical evaluation indicates that the noise adaptive
CNN algorithm and two path architecture make it possible to apply dDLR to clinical
1.5T data with improved image quality. Conclusion
The
noise adaptive dDLR algorithm with two path architecture trained on 3T brain
and knee data can successfully be applied to 1.5T brain and knee clinical
images. Further clinical studies will be
conducted to explore further possible clinical benefits of dDLR. Acknowledgements
No acknowledgement found.References
1. Isogawa K et al. Noise level adaptive deep
convolutional neural network for image denoising. Proceedings of ISMRM Paris,
2018; 2797.
2. Shinoda K et al. Deep Learning Based Adaptive Noise Reduction in
Multi-Contrast MRI; ISMRM 2019 #4701
3 Kidoh M et al. Deep Learning Based Noise Reduction for Brain
MR Imaging: Tests on Phantoms and Healthy Volunteers. Magn Reson Med Sci
doi:10.2463/mrms.mp.2019-0018
4 Cheng JY et al. (2019) Compressed Sensing: From Research to Clinical Practice with Data-Driven Learning (2019).