Hemodynamic Biomarker Profiling in Bicuspid Aortic Valve with 4DFlowNet Deep Learning Models: Post-Processing vs Super-Resolution

Background: Bicuspid aortic valve (BAV) affects ~1–2% of the population [1] and predisposes to aortopathy. Peak-systolic markers—peak velocity (PV), normalized flow displacement (NFD), and wall shear stress (WSS)—are essential for risk stratification, but 4D flow MRI often underestimates them due to noise, artifacts, and limited resolution. We aimed to compare two variants of a common 4DFlowNet-derived backbone for BAV assessment—Enhanced (native-resolution denoising/artifact suppression) and SR (2× isotropic super-resolution of the Enhanced model)—against a reproduced published 4DFlowNet configuration (baseline)[2], and head-to-head (Enhanced vs SR), using internal voxelwise fidelity and external peak-systolic plane-wise PV/NFD/WSS.

Methods: Retrospective IRB-approved 3T 4D flow MRI from 80 BAV patients was analyzed. Seventy cases were for training/validation, with testing on 10 internal and 5 external patients. The internal reference was an expert-processed native-resolution pipeline with semi-automated masks. SR training used a 2× central k-space crop with SNR-matched complex noise, so low-resolution inputs differed from labels mainly in spatial resolution, not SNR.
Internal metrics were mean absolute error (MAE m/s), root mean square error (RMSE m/s), cosine similarity, R², and Bland–Altman vs reference, compared with paired Wilcoxon tests (unadjusted). External metrics at peak systole (per plane) included PV (max through-plane speed m/s), NFD (centroid displacement/radius)[3], and WSS (mean within a fixed 2mm near-wall band Pa). Analyses used per-patient means with 95% CIs (unadjusted), Hedges g, and paired tests with Holm-adjusted p across PV/NFD/WSS. Peak-systolic planes from the native mask were identically applied to baseline/Enhanced/SR to standardize sampling. Plane-wise PV/NFD/WSS were evaluated only in the external cohort.

Results: On the internal test (n=10), vs baseline (MAE 0.113; RMSE 0.206; cosine 0.556), SR improved MAE to 0.082 (−27%) and RMSE to 0.147 (−29%), with (cosine 0.731; R²=0.892). Enhanced further reduced errors (MAE 0.062, −45%; RMSE 0.123, −40%; both p≤0.002 vs baseline), with (cosine 0.845; R²=0.942); Enhanced outperformed SR (MAE 0.062 vs 0.082, p=0.0039; RMSE 0.123 vs 0.147, p=0.037). Enhanced reduced banding, sharpening near-wall gradients. In the external cohort (n=5), SR increased PV vs Enhanced (+1.13; 95% CI 0.905–1.390 ; g=2.12; Holm-adjusted p=0.005; Fig. 1A), while NFD (+0.08; 95% CI 0.03–0.137 ; g=0.81) and WSS (+0.31; 95% CI 0.188–0.454 ;g=1.07) trended higher but were not significant after Holm correction (both Holm-adjusted p=0.127; Fig. 1B–C).

Conclusion: Enhanced achieved higher voxelwise fidelity and artifact suppression, while SR improved jet delineation and PV. NFD/WSS trends were underpowered, and visual assessment confirmed SR’s clearer flow patterns and sharper boundaries (Fig 2), supporting potential utility for early BAV risk stratification. Noise-matched downsampling may understate SR; PV/jet depiction remains clinically salient.

Figure 1. External cohort (n=5), peak systole. Patient-level paired plots comparing Enhanced (blue) vs SR (orange); markers are group means (per-patient mean across planes). Planes fixed at peak systole; WSS from a fixed 2-mm band. PV: Δ mean +1.13 m/s (95% CI 0.905–1.390; Holm-adj p=0.005). NFD: +0.08; WSS: +0.31 Pa (both Holm-adj p=0.127).
Figure 2. Peak-systolic streamlines with identical seeding and visualization parameters comparing Enhanced (left) and 2× super-resolution (SR, right). Colors denote velocity magnitude (0–0.8 m/s). Apparent near-wall texture reflects voxel-sharing/partial-volume at the lumen boundary rather than model-induced fluctuations. Relative to Enhanced, SR concentrates the jet core, sharpens shear layers, and accentuates near-wall gradients.