Purpose To establish a comparability-first cine-MRI paradigm for small intestinal motility using a unified, feature-agnostic normalized differential motion index (NDMI). On this foundation, we investigated scan-rescan consistency and sensitivity under simulated multicenter variability across three methodological families – Optical flow, Morphology, and Implicit neural representation (INR). Methods Twelve healthy volunteers underwent same-scanner scan-rescan cine-MRI. Nineteen high-confidence ROI subsequences were curated. Motility features were derived from (i) generalized Optical flow, (ii) JointSwinUNETR segmentation, and (iii) INR registration adapted to 2D series. A single NDMI parameterization was applied across methods and compared with expert three-level ordinal ratings (vigorous/moderate/low) using accuracy, weighted kappa (xw), and Spearman’s <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\rho \)</EquationSource> </InlineEquation>. Repeatability was assessed with Bland-Altman analysis. Robustness was probed via controlled perturbations (intensity scaling/bias, Gaussian noise, affine distortions), and variance contributions were estimated. Per-frame latency was recorded. Results Agreement with expert ratings was favorable across all families: INR achieved the highest overall accuracy and rank correlation (Acc<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\approx 0.88\)</EquationSource> </InlineEquation>, highest <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\rho \)</EquationSource> </InlineEquation>), Morphology yielded the strongest quadratic-weighted <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\kappa \)</EquationSource> </InlineEquation>, and Optical flow was lower overall, particularly in the moderate class. Signed Bland–Altman analyses showed near-zero bias and the tightest limits of agreement for Morphology with the highest <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\textrm{ICC}\)</EquationSource> </InlineEquation>, intermediate repeatability for INR, and wider limits for Optical flow. Under simulated multicenter perturbations, dominant variance sources diverged–noise plus isotropic scale for Optical flow, geometric distortions for Morphology, and Rician noise for INR–whereas global intensity shifts were negligible within the tested ranges. Mean per-frame latency was 0.03 s (Morphology), 0.07 s (Optical flow), and 5.66 s (INR). Conclusion NDMI provides a unified, feature-agnostic basis for fair cross-family comparison of intestinal motility. Under controlled conditions, all three methods were repeatable: Morphology was most consistent; INR showed the best agreement with expert ratings and relative robustness to affine changes; Optical flow was more sensitive to SNR and scale. These profiles support scenario-aware method selection and motivate prospective multicenter validation with protocol harmonization.

Unified NDMI for small intestinal cine MRI: repeatability and sensitivity to multicenter variability

期刊论文
发表日期：

2025年12月10日

查看全文

Yue Wang,
Wenhui Li,
Wei Wang,
Ziming Wang,
Taoli Du,
Ming Ma,
Mengchao Zhang

摘要

Purpose

To establish a comparability-first cine-MRI paradigm for small intestinal motility using a unified, feature-agnostic normalized differential motion index (NDMI). On this foundation, we investigated scan-rescan consistency and sensitivity under simulated multicenter variability across three methodological families – Optical flow, Morphology, and Implicit neural representation (INR).

Methods

Twelve healthy volunteers underwent same-scanner scan-rescan cine-MRI. Nineteen high-confidence ROI subsequences were curated. Motility features were derived from (i) generalized Optical flow, (ii) JointSwinUNETR segmentation, and (iii) INR registration adapted to 2D series. A single NDMI parameterization was applied across methods and compared with expert three-level ordinal ratings (vigorous/moderate/low) using accuracy, weighted kappa (xw), and Spearman’s \(\rho \) . Repeatability was assessed with Bland-Altman analysis. Robustness was probed via controlled perturbations (intensity scaling/bias, Gaussian noise, affine distortions), and variance contributions were estimated. Per-frame latency was recorded.

Results

Agreement with expert ratings was favorable across all families: INR achieved the highest overall accuracy and rank correlation (Acc \(\approx 0.88\) , highest \(\rho \) ), Morphology yielded the strongest quadratic-weighted \(\kappa \) , and Optical flow was lower overall, particularly in the moderate class. Signed Bland–Altman analyses showed near-zero bias and the tightest limits of agreement for Morphology with the highest \(\textrm{ICC}\) , intermediate repeatability for INR, and wider limits for Optical flow. Under simulated multicenter perturbations, dominant variance sources diverged–noise plus isotropic scale for Optical flow, geometric distortions for Morphology, and Rician noise for INR–whereas global intensity shifts were negligible within the tested ranges. Mean per-frame latency was 0.03 s (Morphology), 0.07 s (Optical flow), and 5.66 s (INR).

Conclusion

NDMI provides a unified, feature-agnostic basis for fair cross-family comparison of intestinal motility. Under controlled conditions, all three methods were repeatable: Morphology was most consistent; INR showed the best agreement with expert ratings and relative robustness to affine changes; Optical flow was more sensitive to SNR and scale. These profiles support scenario-aware method selection and motivate prospective multicenter validation with protocol harmonization.