Encrypted login | home

Program Information

Multi-Layer Spectral Analysis for Tensor Structure Encoding of 4D Deformation Field Data

no image available
A Iliopoulos

AS Iliopoulos1*, Y Zhang2, N Pitsianis3,1, X Sun1, FF Yin2, L Ren2 , (1) Duke University, Durham, NC, (2) Duke University Medical Center, Durham, NC, (3) Aristotle University of Thessaloniki, Greece


WE-D-303-6 (Wednesday, July 15, 2015) 11:00 AM - 12:15 PM Room: 303

Purpose: Adaptive denoising and encoding of 4D deformation-vector-field (DVF) extracted from 4D planning CT, for modeling patient dominant motion features across coupled trajectory dimensions and respiratory phases.

Methods: We propose a multi-layer spectral analysis method to capture inherent structure in 4D deformation fields such as those obtained from phase-correlated 4D-CT data. We illustrate the method with two particular two-layer schemes. One involves phase analysis of 3D trajectories at the first layer, followed by directional analysis of deformation trajectories within each time-spectral component. In the other scheme, we first analyze dominant features in the joint direction-phase trajectory space at the first layer; at the second layer, we separate the directional and temporal variables and discard residual noise which may be correlated in the joint space. With either scheme, directional and phase-wise coherence are preserved. Essentially, we regard the DVF data as a tensor that spans space (voxels), direction (deformation vectors), and time (respiratory phases). The method can be extended to multiple layers and augmented with prior structure. Dominant components across layers may be extracted automatically according to a prescribed dominance percentile.

Results: We have investigated the feasibility and efficacy of the two schemes using a set of DVFs across 10 respiratory phases in a digital XCAT phantom 4D-CT. Dominant multi-layer spectral components are shown to capture a different spectral structure than their single-layer counterparts, attesting to the multi-dimensional nature of motion features. We are proceeding to analyze real patient data where motion patterns and correlations may be more complex and noisy.

Conclusion: The proposed methodology provides an interpretable, adaptable feature space for motion model extraction and compressive encoding. Noise components may be removed in different spectral regions, while preserving temporal, directional, and spatial motion coherence. The resulting motion models have the potential to improve deformable image reconstruction using motion models.

Funding Support, Disclosures, and Conflict of Interest: National Institutes of Health Grant No. R01-CA184173

Contact Email: