Biomarker-Aware Super-Resolution for 4D Flow MRI in a Synthetic Pulmonary-Artery Benchmark
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Purpose
To evaluate, within a synthetic pulmonary-artery benchmark, whether biomarker-aware super-resolution improves hemodynamic endpoint fidelity in 4D flow MRI.
Methods
A 3D residual channel attention network (RCAN) was trained on 22 CFD simulations from 11 pulmonary-artery geometries with 2× spatial degradation. The original primary comparison evaluated full biomarker-aware training against a combined divergence–temporal-regularization control across 14 matched seeds on a fixed six-case test set. Additional matched analyses subsequently separated divergence and adjacent-frame temporal-difference regularization and evaluated biomarker supervision in a controlled 2×2 design with a common reconstruction schedule.
Results
In the original primary comparison, biomarker-aware training reduced PI error relative to the matched combined-regularization control (12.76 ± 2.34% vs. 17.12 ± 1.23%; paired difference −4.35 percentage points [95% CI −5.58, −3.09]; p = 0.0015). In the matched factorial analysis, biomarker supervision reduced PI error without (−2.235 percentage points [95% CI −3.514, −0.849]) and with (−4.125 percentage points [95% CI −5.392, −2.732]) combined regularization. PI improvement was not mirrored by uniformly improved reconstruction metrics: PSNR changed by −0.058 ( p = 0.50) and +0.144 dB ( p = 0.28), while SSIM decreased by 0.009 ( p = 0.004) and 0.007 ( p = 0.009), respectively. Divergence alone changed PI by +0.034 percentage points [95% CI −0.705, 0.809] while reducing DivRMS by 1.869 percentage points; temporal-difference regularization increased PI error by 1.562 percentage points [95% CI 0.594, 2.439].
Conclusion
Within this synthetic, fixed-case RCAN benchmark, biomarker supervision improved PI under both regularization conditions. PI improvement was not mirrored by uniformly improved reconstruction metrics. Temporal-difference—not divergence—regularization was the only isolated component associated with reproducible PI deterioration. These findings are configuration-specific and require geometry-independent and in vivo validation.
