INTRODUCTION

MRI scans often last several minutes during which subject motion can occur. This motion can create image artifacts, such as blurring and ghosting, and reduce the image quality. Depending on the level of image corruption, images might be usable or reacquisition may be required. The main goal of this work was to investigate the capability of different basic machine learning algorithms¹ to classify image quality of simulated and of in-vivo data acquired at ultra-high field MRI.

METHODS

Motion detection with machine learning was tested for three synthesized datasets. The first dataset was created from a 2D Shepp-Logan phantom (256x256 pixels): 10000 images without motion artifacts, and 10000 images containing a known level of motion artifacts, (see Fig.1, column (a)). The second dataset created from a 3D Shepp-Logan phantom (256x256x256 pixels): 12800 images without motion artifacts, 50 volumes with different orientation and position in 3D space and the same number of volumes with motion artifacts, (see Fig.2, column(b)). The last synthetic dataset consisted of 3D T1-weighted (T1-w) brain images, resolution 1.5 mm isotropic, 128x128x128 voxels, 100 volumes (12800 images) were generated without and with motion artifacts (in total 25600 images; see Fig.1, column (c)). All synthesized data included background noise and field inhomogeneities. For the in vivo data, 21 healthy subjects, after written consent, were scanned at 7T (Siemens, Erlangen, Germany) with a 32-channel head coil (Nova Medical, Wilmington, USA), using 3D-MPRAGE with prospective motion correction functionality² (PMC): T1-w images, full brain acquisition, 0.45x0.45x0.45 mm, TR/TE/TI 2820/2.82/1050 ms, FA 5°; and were asked not to move. For tracking rigid body subject motion, a single in-bore camera tracked the position of a marker (MT384ib, Metria Innovation Inc., Milwaukee, WI,USA), which was rigidly attached to the subject’s upper jaw with an individually manufactured mouthpiece. Every subject was scanned with and without PMC, (order of PMC ON – PMC OFF was randomized). Additionally, 3 subjects were asked to follow with their head a determined motion pattern during scanning (done with and without PMC). In total 48 volumes (19968 images, 416 slices for each volume) were acquired, examples are shown in Fig. 2 (c). For all images several features were extracted, for example: mean, variance, standard deviation, texture features⁴ calculated from the gray level co-occurrence matrix (glcm) and features related to the edges (all the data are available at: https://github.com/sarcDV/phantom_data). Support Vector Machine (SVM), Logistic Regression (LR) and Nearest Neighbors (NN) were used to classify the extracted features from the images¹. The synthetic data sets were split randomly into train and test subsets (70%, 30%) for all processing. Also others splitting ratios train/test subsets were tested, (60%, 40%), (50%, 50%).

RESULTS

The synthesized 2D dataset was classified with the SVM algorithm using the image mean and Scharr’s operator³ (edge detector) with 4 different kernels (see Fig. 3). For both 3D synthetic datasets LR was used. This linear model allows to differentiate the data with decision boundaries as shown in Fig. 4 a),b). For the in-vivo dataset LR and NN classification were used. Fig. 5 b) shows the NN classification applied to the group of features: glcm-energy1 (texture analysis⁴) vs. Scharr’s characteristic.

Acknowledgements

This work was supported by the Initial Training Network, HiMR, funded by the FP7 Marie Curie Actions of the European Commission, grant number FP7-PEOPLE-2012-ITN-316716, and by the NIH, grant number 1R01-DA021146.