Synopsis

Motion is still one of the major extrinsic factors degrading image quality. Automated detection of these artifacts is of interest, (i) if suitable prospective or retrospective correction techniques are not available/applicable, (ii) if human experts who judge the achieved quality are not present, or (iii) if a manual quality analysis of large databases from epidemiological cohort studies is impracticable. A convolutional neural network assesses and localizes the motion artifacts. This work extends the previously published method by proposing a general architecture for a whole-body scenario with varying contrast weightings. High accuracies of >90% were achieved in a volunteer study.

Introduction

MRI offers a broad variety of imaging applications with targeted and flexible sequence and reconstruction parametrization. Varying acquisition conditions and long examination times make MRI susceptible to imaging artifacts, amongst which motion is one of the major extrinsic factors deteriorating image quality.

Images recorded for clinical diagnosis are often inspected by a human specialist to determine the achieved image quality which can be a time-demanding and cost-intensive process. If insufficient quality is determined too late, an additional examination may be even required decreasing patient comfort and throughput. Moreover, in the context of large epidemiological cohort studies such as UK Biobank¹ or German National Cohort² the amount and complexity exceed practicability for a manual image quality analysis. Thus, in order to guarantee high data quality, arising artifacts need to be detected as early as possible to seize appropriate countermeasures, as e.g. prospective^3-5 or retrospective correction techniques^6-9. However, these methods are not always available and/or applicable. Moreover, when a human expert is not present or for large cohort studies, the potential presence of motion artifacts demands an automatic processing for a prospective quality assurance or retrospective quality assessment.

Previously proposed approaches for automated medical image quality analysis require the existence of a reference image and/or were only focused on specific sequences and scenarios^10-12. Reference-free approaches^13-17 are mainly metric-based driven to evaluate the quality on a coarse level.

In none of the previous approaches a motion artifact localization and quantification was conducted. We therefore proposed an automatic and reference-free motion artifact detection by a machine-learning convolutional neural network (CNN) approach^18,19. In our previous study, the architecture was kept intentionally shallow because of the limited amount of data and investigations focused only on motion in head and abdominal region for one imaging sequence. In this work, we investigate a general architecture for motion artifact detection in a whole-body scenario with two contrast weightings.

Material and Methods

MR images were acquired on a 3T PET/MR (Biograph mMR, Siemens) from 18 healthy volunteers (3 female, 25±8y) with a T1w and T2w FSE sequence. The acquisition parameters for the respective body regions (head, abdomen, pelvis) are depicted in Tab.1. Each of the five sequences was acquired twice for every volunteer. During the first acquisition volunteers were asked to avoid movements (head, pelvis) or an end-expiratory breath-hold (T1w) respectively navigator-triggering (T2w) was conducted in the abdominal region. In the second acquisition, volunteers were asked to move their heads, hips (rigid deformation) or to breathe normal (non-rigid deformation). Images are normalized into an intensity range of 0 to 1 and partitioned into 50% overlapping patches of size 40x40x10 (APxLRxSI).

The proposed CNN architectures are depicted in Fig.2 and compared against the previously published 2D-CNN¹⁸ to output artifact probabilities p. 3D-CNN depicts the 3D extension of the 2D-CNN with three convolutional layers of N filter kernels/channels of size MxLxB with rectified linear unit (ReLU) activation, followed by a fully-connected dense layer with softmax decision.

The MNetArt is inspired by MNet²⁰ consisting of four stages followed by a dense output layer. Each stage contains two convolutional layers with ReLU activation, an intermittent and finalizing concatenation layer. This resembles a residual path to forward feature maps in and between stages. The first three layers exhibit an additional max-pooling downsampling.

The VNetArt is inspired by VNet²¹ consisting of three stages followed by a dense output layer. Each stage has two convolutional layers with ReLU activation, a finalizing concatenation layer (residual path) and max-pooling downsampling.

The four architectures (2D-CNN, 3D-CNN, MNetArt, VNetArt) are trained by a leave-one-subject-out cross-validation on (i) the complete training database, (ii) the subset consisting only of T1w or T2w images, (iii) subsets leaving-out one body region. Categorical cross-entropy is minimized for given learning rate, $$$\ell_2$$$ regularization and dropout. Parameter ranges are estimated by the Baum-Haussler rule²² with a grid-search optimization. Testing was performed on the left-out subject to determine accuracy, sensitivity (true positive rate; TPR) and specificity (true negative rate; TNR).

Results and Discussion

Conclusion

Acknowledgements

References