Tehran University of Medical Sciences Frontiers in Biomedical Technologies 2345-5837 0 0 2026 10 24 1705 1705 Mohammad Rashidi Department of Biomedical Engineering,SR.C.,Islamic Azad University,Tehran,Iran. Nader Jafarnia dabanloo Department of Biomedical Engineering, SR.C., Islamic Azad University, Tehran, Iran. Fereidoun Nowshiravan Rahatabad Department of Biomedical Engineering, SR.C., Islamic Azad University, Tehran, Iran. Maryam Shojaeifard Echocardiography Research Center, Rajaie Cardiovascular Medical and Research Center, Iran University of Medical Sciences, Tehran, Iran. 2026 08 24 2026 09 06 Purpose: To develop and evaluate a data-driven multi-reference registration strategy for left-ventricular motion tracking in three-dimensional echocardiography, using an intrinsic low-dimensional cardiac trajectory to determine the temporal placement of local registration references. Materials and Methods: Twelve healthy-volunteer three-dimensional echocardiographic sequences from a publicly available myocardial-tracking benchmark were analyzed. We evaluated six low-dimensional representation methods (PCA, t-SNE, Diffusion Maps, LLE, UMAP, and Isomap) in standard and time-embedded forms using geometric and physiologically relevant criteria. Four reference frames were automatically extracted from the resulting two-dimensional trajectories. Time-embedded PCA and Diffusion Maps produced identical four-reference sets for all analyzed cases and therefore resulted in the same proposed registration configuration. We compared five primary registration configurations—single-reference, sequential frame-to-frame, the proposed manifold-guided multi-reference strategy, a partial B-Spline–Fourier (BSF) implementation, and a Fixed Four-Reference Control—using landmark-based Target Registration Error (TRE) in millimeters. We also evaluated an additional Two-Reference Control (K=2) that retained only the initial ED and principal internal references as a reference-number sensitivity analysis. Results: For both time-embedded PCA and Diffusion Maps, the automatically selected principal internal reference coincided with the expert-defined end-systolic (ES) frame in 10 of 12 cases and differed by one frame in the remaining two cases. The proposed strategy achieved mean-cycle, ES-frame, and final-frame TRE values of 6.8 ± 1.6 mm, 11.5 ± 2.7 mm, and 4.0 ± 1.2 mm, respectively. For mean-cycle TRE, Strategy 3 showed a median paired reduction of 0.99 mm relative to the Fixed Four-Reference Control [95% CI: 0.94–1.08; |r_rb|=0.91; p_Holm=0.033]. At ES, the corresponding reduction relative to the Fixed Four-Reference Control was 2.000 mm [1.945–2.055; |r_rb|=1.000; p_Holm=0.021]. The K=2 control showed no significant mean-cycle difference from the other configurations, but Strategy 3 reduced ES-frame TRE relative to K=2 by 1.265 mm [0.304–2.445; |r_rb|=0.909; p_Holm=0.029] and final-frame TRE by 1.156 mm [0.483–2.142; |r_rb|=1.000; p_Holm=0.029]. No statistically significant final-frame difference was detected between Strategy 3 and the Fixed Four-Reference Control. Conclusion: Within this benchmark of 12 healthy-volunteer sequences, the findings support the internal technical feasibility of using the intrinsic low-dimensional organization of a cardiac image sequence for data-driven reference-frame placement within a multi-reference registration architecture. Time-embedded PCA and Diffusion Maps produced identical reference configurations, indicating no additional reference-selection benefit from the more computationally demanding Diffusion Maps representation in the present benchmark. The K=2 sensitivity analysis showed phase-specific effects of removing the intermediate references, while the absence of a significant mean-cycle difference prevents interpretation of K=4 as uniformly superior or empirically optimal. Independent validation in larger and clinically heterogeneous datasets is required before broader generalization. https://fbt.tums.ac.ir/index.php/fbt/article/view/1705 https://fbt.tums.ac.ir/index.php/fbt/article/download/1705/594