Pose-Invariant, Compact Motion Maps for Imaging and Radiotherapy Adaptation
Abstract
Purpose
Accurately modeling how organs move is essential in modern medical imaging and radiotherapy. Tasks such as deformable image registration (DIR), dose accumulation, motion tracking, and treatment adaptation all depend on reliable representations of anatomical deformation. However, widely used motion models based on vector-valued Principal Component Analysis (PCA) tightly couple direction and magnitude at each voxel, making them sensitive to patient pose and requiring many modes to capture respiratory patterns. This complexity limits its clinical usability, increases computational burden, and reduces robustness. To create a robust and computation efficient framework for motion modeling, we develop Anatomical Deformation Modes (ADMs), a compact, pose-invariant organ motion representation that decouples spatial motion pattern (“motion maps”) from vectorized temporal motion coefficients.
Methods
We first analyzed the relationship between the ADM and PCA and demonstrated that ADMs can express PCA model, while the reverse does not hold. Then we evaluated ADMs using simulated deformations and 50 4D lung CT datasets. Reconstruction accuracy, pose invariance, and parameter efficiency were assessed, including reconstruction from limited data such as single slices.
Results
ADMs provided more compact and pose-invariant representations of respiratory motion compared to PCA. In simulated data, ADM spatial maps remained stable under rotations, whereas PCA modes changed significantly with pose. For 4D lung CT, sub-voxel reconstruction accuracy (P95 < 2 mm) was achieved using only two ADM modes, whereas PCA required at least three or more. ADMs demonstrated greater flexibility in selecting the number of modes and higher parameter efficiency while maintaining comparable overall accuracy.
Conclusion
ADMs offer a clinically practical, compact, and pose-invariant framework for modeling anatomical deformation. Their robustness and efficiency make them well suited for a broad range of clinical applications, including fast and stable DIR, motion-aware dose accumulation, real-time motion tracking, and synthetic 4D imaging for plan robustness evaluation.