1351894 - Fast and Multi-Platform Motion Correction of Stress-Perfusion CMR using Principal Component Analysis and Deep Learning

The single-slice pCMR dataset included 34 subjects and two vendors’ scanners - 19 subjects: Philips Achieva 3T, TR/TE 2.22/1.03ms, voxel size 1.25x1.25x10mm. 15 subjects: Siemens Skyra 3T, TR/TE 1.60/1.05, voxel size 0.924x 0.924x10mm. See Figure 3 for example images. Three stages of the cardiac cycle were imaged via ECG signal gating. 

The training data was taken from 26 subjects, while 8 subjects were test cases (4 from each scanner). After applying the bounding box [3], the groupwise motion correction was applied to the training datasets using the existing method [2]. Training deformation fields were generated by combining random affine translations, random complex deformations, simulated minor contractions/expansions of the myocardium, and empirical deformations from the existing method. For each scan, a motionless reference dataset was created using Principal Component Analysis (PCA) in the time dimension (Figure 2), using a regularisation parameter 𝛌= 0.5 / √ (no voxels) which removed motion but retained some contrast features to aid the Convolutional Neural Network (CNN). This produced an augmented dataset of 10,000 moving/reference image pairs with ground truth deformations. The U-Net [4] CNN was developed using PyTorch [5]. The cost function was the total pythagorean vector error of the predicted deformation, plus a total divergence term to penalise discontinuity.

Results: Figure 3 shows the mean intensity curve in the segmented interventricular myocardium for the test subjects, with the effect of motion correction. Without motion correction, the intensity measurement became unstable due to respiratory motion (SD: Siemens 0.923, Philips 0.930). This was improved by the baseline method [2] (SD: Siemens 0.706, Philips 0.734) and by a statistically equivalent amount by the CNN (SD: Siemens: 0.708, Philips 0.725).

Conclusion: We have demonstrated that CNNs can correct pCMR respiratory motion in around 10 seconds, 500x faster than established retrospective method, with similar stabilisation of the tracer-kinetic curve (Figure 3) which is crucial for accurate quantification. In future work the CNN will be trained to handle cases where ECG signal gating fails. The method will be implemented onto the scanner operating system combined with rapid automation of the rest of the pipeline, making the diagnostic power of quantitative pCMR immediate and more widely accessible.