Retrospective motion artifact correction of structural MRI images using deep learning improves the quality of cortical surface reconstructions

Abstract

Head motion during MRI acquisition presents significant challenges for neuroimaging analyses. In this work, we present a retrospective motion correction framework built on a Fourier domain motion simulation model combined with established 3D convolutional neural network (CNN) architectures. Quantitative evaluation metrics were used to validate the method on three separate multi-site datasets. The 3D CNN was trained using motion-free images that were corrupted using simulated artifacts. CNN based correction successfully diminished the severity of artifacts on real motion affected data on a separate test dataset as measured by significant improvements in image quality metrics compared to a minimal motion reference image. On the test set of 13 image pairs, the mean peak signal-to-noise-ratio was improved from 31.7 to 33.3 dB. Furthermore, improvements in cortical surface reconstruction quality were demonstrated using a blinded manual quality assessment on the Parkinson's Progression Markers Initiative (PPMI) dataset. Upon applying the correction algorithm, out of a total of 617 images, the number of quality control failures was reduced from 61 to 38. On this same dataset, we investigated whether motion correction resulted in a more statistically significant relationship between cortical thickness and Parkinson's disease. Before correction, significant cortical thinning was found to be restricted to limited regions within the temporal and frontal lobes. After correction, there was found to be more widespread and significant cortical thinning bilaterally across the temporal lobes and frontal cortex. Our results highlight the utility of image domain motion correction for use in studies with a high prevalence of motion artifacts, such as studies of movement disorders as well as infant and pediatric subjects.

Keywords: Cortical surface; Cortical thickness; Image quality; Motion artifact; Parkinson's disease; T1.

Figure 1:. Summary of the proposed study framework outlining the 5 different stages of model development and testing.

(1) Training the 3D CNN to learn the clean data from motion corrupted data. The model was trained patch-wise in native space using 128×128×128 patches. Preprocessing included intensity normalization and cropping of the input image. (2) Testing on unseen validation dataset using different levels of motion severity and SSIM and pSNR as evaluation metrics. (3) Testing on real motion artifact affected data was carried out using a manual QC evaluation where no ground-truth was available and SSIM/pSNR where a minimal motion paired “ground-truth” was available. (4) Cortical reconstruction quality improvement was assessed using a manual quality control. (5) Examining whether motion correction was able to better identify morphological changes in Parkinson’s disease.

Figure 2:. Schematic of motion artifact simulation along with example images for different simulation parameters.

(a) Schematic of motion artifact simulation, which involves a 3D FFT followed by corruption of lines in the Fourier domain. Phase shifts are used to mimic translational motion and rotations (also in the Fourier domain) to simulate head rotation. (b) Different sampling schemes with image examples. Three sampling schemes were tested: Gaussian, piecewise transient and piecewise constant. For the Gaussian model, artifacts were less coherent compared to the piecewise constant or piecewise transient models where ghosting was more evident. The colorbar indicates the motion magnitude (in voxels) applied in the Fourier domain. (c) Examples of different motion severities for the piecewise constant model, generated by varying the percentage of motion affected Fourier lines. Upper panel – example images. Lower panel – Difference between the corrupted image and ground truth motion free image.

Figure 3:. On the validation dataset CNN models trained with different severity levels (0-20, 0-30 and 0-40) significantly improved both data quality metrics at all levels of motion severity.

(a) Example images before and after correction with the 0-30 model. The corresponding difference image compared to the ground-truth is shown below each image. The ground-truth image is indicated by simulated severity = 0% before correction. (b) Similarity metrics relative to the ground-truth before correction and for different models trained on different severity ranges. Upper panel – structural similarity index (SSIM). Lower panel – pSNR. pSNR is not defined for the before correction 0% corruption case where the images are identical. * represents statistically significant differences at a significance level of p<0.05 (paired t-test). (n=46)

Figure 4:. Motion correction visually and quantitatively improves the image quality of real motion artifact affected data.

Models trained with coherent motion, the piecewise transient and piecewise constant outperformed that trained with samples drawn independently from a Gaussian distribution. Upper panels – Example images. Lower panels – Quality score as assessed manually on a scale between 1-5. (a) Original and corrected results from models trained using different motion simulation approaches. From left-to-right: Gaussian, piecewise transient, piecewise constant, piecewise constant-nufft (with rotations) (b) Model performance on different motion serveries for the piecewise constant model, trained using increasing levels of simulated severity from left-to-right: 0-20, 0-30, 0-40 % of phase-encoding lines. Error bars indicate the standard deviation across n=10 images.

Figure 5:. Example images from five different subjects from the ADNI dataset before and after motion correction compared to a minimal motion reference image.

Left column: minimal motion reference image. Middle column: motion affected image. Right column: Motion affected image after motion correction.

Figure 6:. Motion correction improves the cortical reconstruction quality as indicated by a manual QC procedure.

(a) Proportions and number of images in each category: pass, questionable and fail. (b) Examples of cortical reconstructions that were deemed to be of questionable quality before motion correction but passed QC after correcting the image and re-running the reconstruction pipeline. (c) Examples of cortical reconstructions that failed QC before motion correction but passed QC after correcting the image and re-running the reconstruction pipeline. The orange arrowheads indicate specific areas of QC failures.

Figure 7:. Cortical thinning in Parkinson’s disease is more significant and more widespread after motion correction compared to before correction.

Left column: T-statistic maps at FDR corrected p<0.05 threshold, indicating significant differences in cortical thickness between PD subjects and controls. Blue regions indicate significantly reduced thickness. Right column: Cortical thickness (adjusted for age and sex) vs. group for the 8 most significantly different regions identified after motion correction. P-values are shown before correction for multiple comparisons. (a) Before applying motion correction, decreases in cortical thickness were limited to the orbital frontal cortices, the left inferior frontal gyrus, middle and superior temporal gyri, the right temporal pole, right insula as well as the left and right anterior cingulate and parahippocampal gyri. (b) After motion correction, there were more significant and more widespread decreases in cortical thickness across both temporal lobes as well as the superior frontal gyrus. (CTL: n=247, PD: n=309)

Figure 8:. T-statistic maps and box plots indicating regions of significant cortical thinning in PD vs. Controls upon only including cortical reconstructions that passed QC i.e. after excluding questionable surfaces in addition to those that failed QC.

Left column: T-statistic maps at FDR corrected p<0.05 threshold, indicating significant differences in cortical thickness between PD subjects and controls. Blue regions indicate significantly reduced thickness. Right column: Cortical thickness (adjusted for age and sex) vs. group for the 8 most significantly different regions identified after motion correction. P-values are shown before correction for multiple comparisons. (CTL: n=238, PD: n=300)