QF1-011 - Improved Phase Unwrapping for DENSE MRI using Deep Learning

Background: DENSE myocardial strain imaging is a method wherein tissue displacement is encoded in the image phase. Phase wrapping is common in DENSE, and phase unwrapping is a necessary step in the strain calculation. A widely-used phase unwrapping method for DENSE based on path-following [1] benefits from user-defined seed points and has a high failure rate when the signal-to-noise ratio (SNR) is low. We propose a deep-learning (DL) approach for DENSE phase unwrapping to eliminate user intervention and improve performance for low-SNR data.

Methods: We used DENSE data from 74 subjects (a mix of healthy subjects and patients) acquired using 1.5T or 3T scanners (Siemens). Each subject included 1-5 short-axis slices with 20-59 cardiac phases. Ten subjects including healthy volunteers and patients imaged at both field strengths were selected for the test data (2,416 total images). Training data included 12,415 short-axis DENSE images from the remaining 64 subjects. We defined the phase unwrapping problem as a semantic segmentation problem [2]. As phase wrapping within the myocardium in DENSE is nearly always confined to one cycle, the semantic segmentation classes were no wrap, –π wrap, or +π wrap (Fig. 1). We assume that, prior to phase unwrapping, anatomical myocardial segmentation has been performed. To create ground truth phase-unwrapped images, we used the path-following method and discarded all frames with visually-detectable unwrapping errors. A U-Net with dilated convolutions in the contracting path was trained with a pixel-wise cross-entropy loss function and an Adam optimizer with a learning rate of 5E-4 and a mini batch size of 10 for 200 epochs. To improve network performance and avoid overfitting, we applied data augmentation by adding noise, performing affine deformation and manipulating the unwrapped ground truth data to generate new wrapped data. For the test set of 10 subjects, we used the DICE coefficient to compare the U-Net and the ground truth. To assess the U-Net for low-SNR data, we added noise to the original test data and we compared the U-Net with the path-following method using these data. For the low-SNR dataset, we calculated the mean squared error (MSE) between the unwrapped phase and the ground truth.

Results: For the test set of 10 subjects, the Dice coefficients were 99.9%, 93.0%, and 92.9% for pixels without wrap, with –π wrap, and with +π wrap, respectively vs. ground truth. Figure 2 illustrates an example where the U-Net and the path-following method were both successful for high-SNR data, but only the U-Net was successful for low-SNR data. For the low-SNR test set, MSE values were 0.05 and 0.21 for the U-Net and the path-following method, respectively. Confusion matrices for the U-Net and the path-following method for the low-SNR test data are reported in Table 1.

Conclusion: A semantic-segmentation U-Net eliminates the need to manually identify seed points and reduces MSE by 75% for low-SNR data.

This work was supported by R01HL147104.