2450 - Trajectory Aware Multiview Analysis of Functional Liver MRI Using Dynamic Time Warping and Multidimensional Projections

Purpose/Objective(s): Early identification of heterogeneous functional liver response remains challenging but may be important for clinical outcomes. Patients with impaired liver function may be more susceptible to severe radiation-induced toxicity. We hypothesized that a dynamic time warping (DTW)–based multiview analytics tool applied to longitudinal 0.35 T MRI biomarkers could detect clinically meaningful similarity patterns and reveal patient subgroups of liver functionality.

Materials/Methods: We developed an interactive visual analytics platform integrating DTW similarity, hierarchical clustering, multidimensional scaling (MDS), t-distributed stochastic neighbor embedding (t-SNE), k-nearest neighbor (kNN) network visualization, and parallel coordinate analysis. The cohort included N = 20 patients with longitudinal functional liver metrics derived from 0.35 T MRI. For heatmap analysis, pairwise DTW distances were computed using Normalized CNR All trajectories to provide a single-biomarker similarity view, with patients hierarchically ordered to generate a strict lower-triangular heatmap. For multivariate t-SNE and network analyses, DTW distances were computed using combined trajectories of Time, Normalized CNR All, SNR Tumor, and SNR Liver to enable higher-dimensional characterization. Parallel coordinates displayed Age, Treated/Liver percentage, and Diagnosis, with lines colored by treatment burden. The primary endpoint was identification of coherent similarity subgroups.

Results: The framework revealed structured heterogeneity in functional liver response. Functional differences in liver were quantitatively captured by the DTW distance metric. A compact subgroup (Patients 13, 16, 5, 10, 6, and 11) showed consistently low mutual DTW distances, indicating highly similar temporal liver function trajectories, while demonstrating larger distances relative to patients 20, 12, 15, 8, 2, and 4, suggesting a distinct functional phenotype. Patient 20 (Y90-treated) separated from the tightly grouped cohort, demonstrating sensitivity to clinically meaningful trajectory differences. MDS, t-SNE, and kNN networks showed concordant neighborhood structure, supporting robustness. Parallel coordinate analysis demonstrated variability in treated liver fraction across clusters, providing clinical interpretability.

Conclusion: We developed an interactive DTW-based multiview analytics platform that identifies meaningful patient similarity structure from longitudinal functional liver MRI. While this exploratory cohort was not powered to correlate imaging phenotypes with toxicity or disease control, by aligning patient trajectories with previously observed functional response patterns, the DTW-based framework may enable treatment personalization encompassing patient-specific risk, inform adaptive therapy, and support biomarker-driven stratification pending validation in larger cohorts.