Title: Mapping Definitional Discrepancies in High-Dimensional Data via Manifold Alignment: A Case Study in 3D Right Ventricular Strain

Abstract: In medical imaging research, it is standard practice to utilize a single sample per participant under the assumption that this snapshot adequately captures their physiological characteristics. However, the way input descriptors are defined or calculated—often stemming from a lack of field-wide consensus—can significantly influence analytical outcomes, yet these variations are frequently overlooked. This study introduces a novel approach leveraging representation learning to generate a parametric map that quantifies the effect of such definitional inconsistencies on a specific physiological descriptor derived from medical imagery. We treat these varying definitions or computational methods as distinct high-dimensional datasets, which may be of diverse nature. Our primary focus is on myocardial deformation (strain), a metric for which there is currently little standardization. To address this, we employ manifold alignment to synchronize the latent representations corresponding to the various definitions of the descriptor. Subsequently, we construct plausible distributions within the latent space to depict the divergence caused by these definitional choices, allowing for the reconstruction of a high-dimensional parametric map that visualizes this divergence. Given the absence of a reliable ground truth for this clinical scenario, we initially validate the method using synthetic datasets before applying it to right ventricular strain data. This latter analysis utilizes 3D echocardiographic sequences from subjects, incorporating multiple strain types available at each point on the endocardial surface mesh of the right ventricle. While this serves as an illustrative example, the proposed methodology holds promise for broader application in population studies involving heterogeneous high-dimensional descriptors.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC