Purpose

MRI is a versatile imaging modality which allows the precise and non-invasive assessment of anatomical structures and pathophysiological pathways. A flexible sequence and reconstruction parametrization make it very tunable to specific applications to meet the demanded diagnostic requirements. These quality criteria and the respective parametrizations need to be optimized and ensured to provide sufficient quality for changing acquisition conditions often without the availability of reference data. The relatively long examination times and varying acquisition conditions predispose it to imaging artifacts which may result in a deterioration of diagnostic image quality. Artifacts can originate from hardware imperfections, signal processing inaccuracies and, in most cases, from the compliance and behavior of the patient which can demand a re-scan if the image is too severely affected by e.g. motion. However, the quality with respect to the diagnostic application is often just checked retrospectively by manual reading of a human observer (HO) when the patient has already left the facility, i.e. a re-scan is just possible with larger effort in an additional exam. Careful operation together with prospective or retrospective correction techniques^1,2 may reduce artifact burden, but are either not always applicable and/or cannot correct for all different types of induced artifacts. It is therefore desired to have an automatic image quality assessment directly after the acquisition with feedback of the acquired image quality and a suitable sequence re-parametrization suggestion to the operating technician. We therefore proposed a non-reference MR image quality assessment (IQA) system based on a machine-learning approach to predict HO labeling scores of arbitrary input images corrupted with unknown artifacts^3,4,5. A reliable quality prediction depends on the derived features. In this work, we will therefore focus on the automatic detection and quantification of motion artifact examples of rigid head motion and non-rigid respiratory motion in the thorax. Motion detection methods aim to track the underlying motion by an MR navigator^6,7,8 via external sensors^9,10,11 or from the acquired time-series images, e.g. via registration^12,13. But no motion quantification or localization has been conducted up to now. We propose the usage of a convolutional neural network (CNN) to quantify and localize motion artifacts for IQA.

Material and Methods

The proposed CNN is shown in Fig.1. The CNN consists of three convolutional layers for the detection and localization of the artifacts. A subsequent fully connected multilayer perceptron with a sigmoid activation function feeds two soft-max output nodes for a probabilistic quantification p of the artifact burden. Input images are patched into sizes of 40x40 overlapping patches to enable the localization and to reduce the computational complexity respectively the number of parameters to be trained. Each convolutional layer consists of N filter kernels with size MxLxB, a rectifying linear unit activation function and a downsampling layer. A dyadic increase in the filter kernel depth N provides a multi-resolution approach. Filter coefficients and bias values are trained via a steepest gradient descent. Parameter optimization and training is conducted by means of a 10-fold cross validation on 10 volunteer datasets (3 female, age 27±4.5 years). Datasets are acquired for each volunteer on a whole-body 3T PET/MR (Biograph mMR, Siemens) with a T1w-TSE in the head and a T1w-GRE in the abdomen. For each region, two datasets are acquired with and without motion artifacts (head: rigid head tilting, abdomen: non-rigid respiratory movement) which serve directly as labels for the supervised training.

Results and Discussion

Fig.2 and 3 display the acquired images of one subject with and without motion artifacts in the head and abdomen, respectively, with a colored overlay of the probability map p for the patches. It can be seen that the artifact affected regions are correctly localized and show a high probability of motion artifacts. For all subjects an average test accuracy of 85% in the head and 70% in the abdomen was achieved. A detection of motion is thus possible in both regions with a superior performance in the head region, because of the more systematic artifact manifestation in the image. These initial results indicate the potential feasibility, but the study has its limitations. So far no patch-based labeling (motion-free and motion-affected) was conducted and the artifact detection was just targeted on motion artifacts on a specific sequence of a single anatomic region.

Conclusion

We propose an algorithm to detect, localize and quantify motion artifacts in the head and abdominal region for a T1w sequences. Future studies will investigate the reliability and robustness to transfer the results to other sequences, body regions and artifacts. This would enable the usage as a reliable automatic predictor in the context of an automatic IQA.

Acknowledgements

No acknowledgement found.

References

[1] McClelland et al., Med Image Anal 2013:17(1). [2] Godenschweger et al., Phys Med Biol 2016:61(5). [3] Küstner et al., Proc ISMRM 2015. [4] Küstner et al., Proc ISMRM 2016. [5] Liebgott et al., Proc ICASSP 2016. [6] Ehman and Felmlee, Radiology 1989:173. [7] Fu et al., MRM 1995:34. [8] Ward et al., MRM 2000:43. [9] Derbyshire et al., JMRI 1998:8. [10] Zaitsev et al., Neuroimage 2006:31. [11] Maclaren et al., PLoS One 2012:7. [12] Kim et al., IEEE TMI 2010:29. [13] Metz et al., Med Image Anal 2011:15.