Purpose

Brain MRI has become a part of the clinical practice to provide diagnostic insights into structural changes associated with neurodegenerative diseases¹. Given the close relationship between neurodegenerative changes and clinical symptoms, the volumetric measurement on structural MRI serves as a potent biomarker for these conditions². To facilitate precise assessments, it is imperative to achieve less noisy and higher contrast imaging. Unfortunately, conventional image reconstruction methods face inherent limitations, as they typically have the trade-off between signal-to-noise (SNR) ratio and scan duration, making it challenging to achieve superior image quality within a relatively short acquisition time. Recently, deep learning-based image reconstruction (DLR) has been commercially introduced to improve the image quality through denoising using deep convolutional neural network (DCNN), thereby allowing the acquisition of high-quality images in a shorter time³. Therefore, we investigated the practical usefulness of DLR in structural MRI for brain volumetry by comparing images reconstructed using conventional methods with those reconstructed using deep learning techniques.

Materials & Methods

Brain structural MRI data were acquired from ten healthy subjects (M/F=6/4; 25~50 years old) on 3T MRI (Canon Medical Systems, Vantage Galan STD) equipped with a 16-channel head and neck coil. From each subject, three different T₁-weighted structural MR images were obtained using a 3D-MPRAGE sequence (TI/TR/TE=1000/7.5/3.3ms, flip angle=10°, spatial resolution=1×1×1mm³, SPEEDER (SENSE) acceleration along the PE direction=2); 1) a original image without any filtering (NEX=1, scan time=5 mins 58 secs), 2) an image with DLR (Advanced intelligent Clear-IQ Engine, AiCE), and 3) a high SNR image assumed as a potential ground truth dataset (NEX=3, scan time=17 mins 54 secs). The DLR approach we adapted is based on a DCNN trained with extensive pairs of low and high SNR images³. This network discriminates between signal and noise, allowing the removal of noise contribution from the original images. It effectively reduces noise, particularly in central part of the images, addressing non-uniform noise patterns by the g-factor in parallel imaging. For brain voxel-based morphometry (VBM), we performed the preprocessing pipeline using CAT 12⁴ toolbox integrated into the SPM12 software⁵. The images were spatially normalized to the standard MNI space and segmented into tissue classes of gray matter (GM), white matter, and cerebrospinal fluid using the DARTEL approach, resulting in a cubic resolution of 1.5 mm. The normalized GM maps were subsequently smoothed with a 6-mm FWHM Gaussian kernel. Additionally, we obtained volumetric parcellations describing subcortical GM volumes of six ROIs including caudate, putamen, pallidum, thalamus, hippocampus, and amygdala labeled by the Neuromorphometrics atlas (http://www.neuromorphometrics.com/). To investigate potential volumetric differences in GM maps across the entire brain between groups, two-sample t-test (e.g., original vs. DLR, original vs. NEX=3) was performed with statistical significance set at an uncorrected p < 0.001. For quantification of subcortical areas, repeated one-way ANOVA was conducted, followed by Bonferroni multiple comparison test, with significance at p < 0.05.

Results

Figure 1 shows a comparison of axial images and SNR maps without and with DLR approaches. In subtraction of images with DLR from original images, the non-uniform noise was reduced without significant loss of structural information (Fig.1A-iii). In VBM analysis, GM volumes obtained from original images (NEX=1) commonly showed a significant reduction in bilateral caudate, putamen, hippocampus, and left amygdala regions, compared with those derived from higher SNR images (NEX=3, Fig.2A) and images with DLR (Fig.2B). No significant alterations were observed in the reverse contrast or between the higher SNR images and images with DLR. Furthermore, the comparison of subcortical volumetry across different image acquisitions (Fig.3A) exhibited volume variability similar with the brain VBM analysis. In general, the volumes measured in original images were underestimated compared to images with higher SNR, except for the pallidum, where an overestimation was observed. These findings from both VBM and subcortical volumetry underscore the substantial effects of image SNR on the quantitative measures of automated brain morphometry.

Discussion & Conclusion

The quantitative analysis of brain MRI is increasingly recognized as valuable complement tool in the clinical practice⁶. Using T₁-weighted brain MRI with open-source software package for automated brain volumetric measurement, we demonstrated that the level of image noise directly impacts on performances of tissue segmentation and quantitative assessments of brain volumetric changes. The brain atrophy in basal ganglia and hippocampal regions, which exhibited volumetric differences based on image qualities in this study, are well-established as clinical imaging biomarkers associated with neurodegenerative diseases such as Alzheimer's disease² and frontotemporal dementia⁷. Therefore, our findings suggest the potential advantages of denoising approaches with DLR, not requiring additional scan time, to enhance the reliability of brain volumetric quantification as screening tools in clinical decision-making.

Acknowledgements

No acknowledgement found.