On a synthetic pulmonary-artery benchmark, biomarker supervision cut 4D flow MRI pulsatility-index error from 17.12% to 12.76%

Biomarker-aware training reduced hemodynamic endpoint error: in the original primary comparison, full biomarker-aware training lowered PI error from 17.12 +/- 1.23% to 12.76 +/- 2.34% relative to the matched combined divergence-temporal-regularization control, a paired difference of -4.35 percentage points (95% CI -5.58, -3.09; p = 0.0015). Prior super-resolution evaluations often focus on image reconstruction metrics; this work puts a biomarker (PI) directly into the training objective and compares against a matched control across 14 matched seeds on a fixed six-case test set, shifting evaluation from pixel similarity toward hemodynamic endpoint fidelity. Paired statistics with 95% confidence intervals and p-values on 22 CFD simulations from 11 pulmonary-artery geometries, 2x spatial degradation, 14 matched seeds, and a fixed six-case test set; all data are synthetic, and the authors state the findings are configuration-specific and require geometry-independent and in vivo validation.

In the controlled 2x2 factorial analysis, biomarker supervision reduced PI error by 2.235 percentage points (95% CI -3.514, -0.849) without combined regularization and by 4.125 percentage points (95% CI -5.392, -2.732) with it. By separating divergence regularization from adjacent-frame temporal-difference regularization and using a 2x2 design with a common reconstruction schedule, the work disentangles training components that had previously been combined, allowing the biomarker-supervision effect to be estimated under each regularization condition. Matched factorial design with a common reconstruction schedule; confidence intervals exclude zero in both conditions, though the sample remains a synthetic benchmark with a fixed test set.

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). This indicates that optimizing for a hemodynamic endpoint and conventional image-quality metrics can diverge, providing a concrete contrast for super-resolution evaluations that report both task and reconstruction metrics. Paired changes, p-values, and replication across two regularization conditions are reported; the SSIM decreases are statistically significant but small and measured only on the synthetic benchmark.

Among isolated components, divergence regularization changed PI by +0.034 percentage points (95% CI -0.705, 0.809) while reducing DivRMS by 1.869 percentage points, whereas temporal-difference regularization increased PI error by 1.562 percentage points (95% CI 0.594, 2.439). The work attributes the combined-regularization effect to specific components, identifying temporal-difference rather than divergence regularization as the only isolated component associated with reproducible PI deterioration, offering a testable distinction for future regularization design. Based on component separation within the matched analysis with confidence intervals; the divergence-regularization PI interval crosses zero, the temporal-difference interval does not, and both come from the same synthetic benchmark.

The results are aimed at researchers developing 4D flow MRI super-resolution on a synthetic pulmonary-artery benchmark, under an RCAN backbone, 2x spatial degradation, a fixed six-case test set, and matched random seeds. They enable follow-up work to add biomarker supervision explicitly to the training objective and to treat divergence and temporal-difference regularization as separable design variables rather than a single combined regularizer. For clinical or product-oriented readers, the evidence currently supports method choices only within the synthetic benchmark and does not yet constitute a performance promise in vivo or across geometries.

The authors explicitly state that the findings are configuration-specific and require geometry-independent and in vivo validation, so whether the PI improvement persists on real 4D flow MRI data, different pulmonary-artery geometries, and different degradation levels remains open. PI improvement coexisting with SSIM decreases means the choice of model-selection metric still needs weighing; divergence regularization reduced DivRMS without significantly changing PI, and its clinical meaning remains to be defined. In addition, this reading is incomplete, covering only the abstract and the competing-interest statement, so method details, figures, and supplementary analyses were not included, and more specific judgments about network architecture, training details, and statistical models cannot be given here.