Fig 1.
Segmentation and mesh-preprocessing pipeline.
The aorta segmentation of each subject is re-meshed and smoothed in an automatic pipeline. This is followed by clipping of the inlets/outlets and head & neck vessels.
Fig 2.
An example aortic shape approximation using our SSM is shown. Individual surface or volumes can be reconstructed using a mean aortic shape and applied deformation field initialised on a set of control points (n = 172).
Fig 3.
Surfaces have flow extensions added before volume meshing. The same simulation set-up is applied to the final volume mesh for each synthetic subject.
Fig 4.
CFD results are interpolated onto a point-correspondence mesh (generated by the SSM), thus restoring node concordance.
Fig 5.
The general sequential, fully-connected DNN set-up used to build both pressure and velocity predictors (’CFD vector’ can be either pressure or velocity PCA vectors).
Fig 6.
CFD gradient extraction method.
Using subject centrelines, 99 plane-averaged pressure or velocity points along the length of the aorta are extracted by sampling the 3D flow fields. The origin is always the aortic root (excluding the extension).
Fig 7.
Left: first three modes of deformation from the SSM (SD = standard deviation). Right: examples of synthetic post-CoA aortas from the test set (using combinations of all 35 shape modes).
Table 1.
Comparison of aorta dimensions between original real cohort (n = 67) and synthetic cohort (n = 3000).
Fig 8.
Left: Distribution of mean nodal errors (MNAEN), computed on the test set (n = 200). Errors are absolute values and are projected on the template aorta. Right: Bland-Altman plots for the overall aorta. Normalised error (%) refers to the NAE of each node in every test case, without taking the absolute value (n = 5,800,000). Only 1,000 randomly selected points were drawn to improve graph readability.
Fig 9.
Best, median and worst test-set predictions.
Comparisons between ground truth (CFD) and predicted (ML) in the test set (n = 200). Best, median and worst cases for both pressure and velocity-magnitude are shown, ranked using the mean node-to-node error (MNAES). Pressure/velocity gradients are also displayed (black lines = CFD, red = ML).
Fig 10.
Scatter plot comparing the shape PCA mode values against subject error (MNAES) in the test set (n = 200). Pearson R coefficients and p-values were computed for each subplot.
Fig 11.
Best and worst pressure and velocity predictions on the real patient test cohort (n = 10). FD (SSM) is the error between the predicted (red) and SSM CFD (dashed black) gradients. FD (real) is the error between the predicted (red) and true CFD (solid black) gradients.