Introduction

Magnetic Resonance Elastography (MRE) is a technique which delivers acoustic waves to the patient and images them with a phase-contrast sequence to calculate quantitative stiffness of tissues.[1] MRE has excellent accuracy for predicting liver fibrosis, and is used clinically by over 1000 sites. Analysis of MRE images requires dedicated training to avoid unique stiffness-related artifacts, and is subjective. Even experienced readers take 15-20 minutes to process a 4-slice MRE exam, and calculate stiffness values with a coefficient of variation of 8-10%.[2] An automated MRE analysis method (ALEC) has previously been proposed, and is currently used to analyze >1000 MRE exams annually at our institution.[3] However, the workflow involves an MRE reader verifying every automated ROI before the results are reported clinically. To reduce the automated ROI failure rate and, consequently, the burden of manual verification/modification, we propose a new tool based on a multi-channel fully-convolutional neural network architecture (CNN).
An MRE dataset consists of acquired magnitude and phase images, and derived stiffness and confidence images (Figure 1). ROI selection in MRE is a multi-channel segmentation problem aiming to select tissue within the liver without inclusions (e.g. tumors, major vessels) or stiffness-affecting artifacts. An ROI covering all valid liver area leads to best reproducibility. Human readers draw ROIs of varying sizes, however, and two entirely non-overlapping ROIs may both be valid, which poses a problem for training a neural network. Similarly, while automated ROI failures are easily detected and modified, there is substantial variation in whether and how modification is done for sub-optimal ROIs. The validation in this study focused on improving the clinical workflow by reducing the number of failures (incorrect stiffness and/or major anatomical leak) and sub-optimal ROIs (smaller non-stiffness affecting leak or small ROI area).

Methods

Training. 4600 consecutive clinical MRE exams were retrieved under a consent-waived IRB from the clinical workflow. Automated ROIs produced by the old method were retrospectively reviewed by the same experienced reader and modified when necessary to remove stiffness artifact, anatomical leakage of >1% of ROI area, or to increase ROI size if it covered <50% of useable liver tissue. The manually reviewed/modified ROIs were used as the training target. Multi-channel data for each imaging slice was treated as an independent training example. Data was split into a 3600 exam training set (14400 examples), and 1000 exam test set.
Algorithm. Each example was centered and cropped down to the patient size, with coordinates retained to inverse-transform the ROI. Images were downsampled to 64x64, and a 12-layer 3-channel fully convolutional network was trained on the magnitude, stiffness, and confidence images to calculate ROIs by optimizing dice loss.
Evaluation. Dice coefficient and stiffness agreement with the target ROIs were calculated. Based on the MRE reproducibility threshold of 19%, ROIs with a stiffness difference of >=20% with the target were considered failed. Anatomical ROI leakage was assessed visually and cases with >10% leak were labeled as failed and 1-10% as sub-optimal. Leak of <1% area into partial volume was considered acceptable, as these cases would not be modified clinically. ROIs covering <50% of useable area were considered sub-optimal. Sub-optimal ROIs are cases likely to be modified clinically but do not carry a risk of wrong diagnosis if the reader fails to modify them as the reported stiffness is within MRE reproducibility limits.

Results

The dice coefficient between CNN ROIs and target ROIs reached 0.87 in the training set and 0.86 in the test set. The difference in calculated stiffness was 0.94% +/- 3.25% (mean +/- SD) in the training set and 1.37% +/- 7.54% in the test set, with correlation coefficient being excellent (>0.98) in both sets (Figure 2). The proposed CNN method improved the failure rate from 5.7% for the old method to 3%, and the rate of sub-optimal ROIs from 10.7% to 2.5%. The breakdown of types of issues for the two methods is shown in Figure 3. Examples of CNN improving on ALEC by avoiding ROI leakage into ascites and avoiding a vessel-related soft spot are shown in Figures 4 and 5. CNN analysis time for a single exam on an office PC was <5 seconds while ALEC takes around 3 minutes.

Discussion

This study uses categorical validation of ROIs with the aim of evaluating how the tool will perform in the clinic. The stiffness difference of the CNN with the target was better than agreement between experienced human readers.[2-3] A comparison of stiffness agreement with target ROIs for the CNN and ALEC is, unfortunately, impractical as most of the target ROIs come directly from ALEC. The failure rates observed in this dataset are higher than those seen in validations of ALEC. We believe this to be a consequence of having a more challenging dataset resulting from pulling all series that were transferred to the clinical workstation. A spot check showed that some of the most difficult examples had been considered failed exams, and a different series was used in the clinical report.

Conclusions

The CNN-based MRE analysis method substantially reduces the number of automated ROI failures, and has an excellent agreement with the final results of the reader-verified workflow.

Acknowledgements

This work was supported by NIH EB001981.

References

1. Muthupillai R, Ehman RL. Magnetic resonance elastography. Nat Med. 1996;2(5):601-3. 2. Serai, SD, et al. Repeatability of MR Elastography of Liver: A Meta-Analysis. Radiology. 2017. 285(1):92-100. 3. Dzyubak B, Venkatesh S, Manduca A, Glaser J, Ehman R. Automated liver elasticity calculation for MR elastography. J Magn Reson Imag. 2016;43(5):1055-63.