1343939 - Deep learning flow imaging without velocity encoding

In phase contrast flow imaging (1), velocity encoding is used to generate a phase change that varies linearly with flow. To measure the flow-induced phase change, both images with and without velocity encoding must be collected. This slows down data acquisition, posing a limitation on temporal resolution of flow measurements. The presented work aims to develop a deep learning approach to measuring flow only from images collected without velocity encoding for improved temporal resolution in flow imaging.

Methods: Flow always affects MR signals. However, the flow-induced variation without velocity encoding has nonlinearity that cannot be assessed with traditional methods. The presented work sought to use deep learning to calibrate the nonlinear relationship between flow and MR signal variation, making it possible to measure flow from images collected without velocity encoding. For feasibility demonstration, a deep learning model (Figure 1) was developed with convolutional and fully-connected neural networks (2, 3). We retrospectively selected 271 human subjects who had undergone through-plane phase contrast flow imaging in the aorta and the pulmonary artery. A set of training data was generated from 255 subjects and a set of testing data from 16 subjects. During training, the images collected without velocity encoding were used as the model input. The phase difference maps between the images collected with and without velocity encoding were used as the label. The training was run in 1000 epochs on an NVIDIA TESLA V100GPU (32GB memory). During testing, only the images without velocity encoding were used. We investigated whether the trained model would generate velocity maps that agreed with those measured from both images with and without velocity encoding. To demonstrate resolution improvement, the training data were manually undersampled such that the testing data had a temporal resolution two times higher than that of the training data.

Results:

Figure 2 shows a set of images collected without velocity encoding and the resultant deep learning velocity maps in reference to the phase contrast velocity maps. The images presented considerable variation associated with flow. The deep learning maps gave flow patterns similar to those in the phase contrast maps. Figure 3 (a) provides an example of the flow measured with deep learning in reference to that from phase contrast flow imaging. They agreed well with each other. Figure 3(b) shows that, in all the testing subjects, the flow measurements with deep learning were strongly correlated (R≥0.8, P< 0.001) to those with conventional methods that use both images with and without velocity encoding.

Conclusion:

We have developed a deep learning approach to measuring flow without velocity encoding and demonstrated its feasibility in 1D flow imaging. The presented work indicated that deep learning would offer the potential to accelerate flow imaging for improved temporal resolution.