TL;DR
This paper introduces a simplified, more efficient pipeline for correcting motion in dual-gated cardiac PET scans. By decoupling respiratory and cardiac motion into separate, individually corrected components, they significantly reduce computational complexity and improve image quality, making advanced motion correction more accessible.
What Problem Does This Paper Address?
Positron Emission Tomography (PET) is a powerful imaging technique, but its long acquisition times (minutes) lead to significant image degradation due to patient motion, particularly cardiac and respiratory movements. This motion causes spatial blurring and inaccuracies in attenuation correction, compromising diagnostic accuracy. While gating (dividing data by breathing/heart phases) helps, it increases noise. Previous methods for correcting both types of motion were computationally intensive, requiring numerous individual registrations. This paper tackles the challenge of making dual-gated cardiac PET motion correction more computationally feasible and robust.
Methods
The core innovation is a simplified pipeline that exploits the assumption of independent cardiac and respiratory motion. Instead of treating the combined motion as a single complex transformation, they split it into two distinct components:
- Respiratory Motion Correction: First, they estimate and correct for respiratory motion. To stabilize this process, they average the cardiac phases within each respiratory phase to create a "cardiac average" image. This averaging reduces noise and improves the reliability of the respiratory motion estimation.
- Cardiac Motion Correction: After respiratory motion is compensated, they then estimate and correct for cardiac motion. Similar to the respiratory step, they average the respiratory phases within each cardiac phase to create a "respiratory average" image, which is then used to estimate cardiac displacements.
This sequential approach significantly reduces the number of registration problems. For a 5x5 gating scheme (5 respiratory and 5 cardiac phases), the number of registration tasks drops from 24 to 8. They employ a mass-preserving hyperelastic registration method, which is crucial for PET data to account for intensity variations due to partial volume effects and to ensure diffeomorphic transformations (guaranteeing invertibility and preventing volume collapse/expansion).
Datasets
The study evaluated their method on data from 21 human patients (19 male, 2 female; 37-76 years old) with known coronary artery disease. The PET data was acquired using a Siemens Biograph Sensation 16 PET/CT scanner in list mode. For the study, the data was cropped to the first three minutes of acquisition. Dual gating into 5 respiratory and 5 cardiac phases was performed. The images were reconstructed using the 3D EM software EMRECON and sampled at 128x100x44 voxels.
Key Insights and Ideas
- Decoupling Motion: The key insight is that separating respiratory and cardiac motion into independent components, rather than treating them as a single complex transformation, drastically simplifies the registration pipeline.
- Averaging for Stability: Averaging across cardiac phases for respiratory motion estimation (and vice versa) is a simple yet effective strategy to stabilize the registration process and reduce noise, especially in dual-gated schemes.
- Computational Efficiency: The reduction in registration subproblems leads to a significant speedup, making the method more practical for clinical use and allowing for finer gating schemes.
- Mass Preservation is Crucial: The use of mass-preserving registration is highlighted as essential for accurate motion correction in PET, ensuring that the total signal intensity is conserved despite the transformations.
Key Results
- Reduced Registration Problems: The proposed pipeline reduces the number of registration tasks from $m \times n - 1$ to $m + n - 2$ for $m$ cardiac and $n$ respiratory gates. For a 5x5 gating, this is a reduction from 24 to 8 registration subproblems.
- Improved Image Similarity: Normalized cross-correlation (NCC) between gated images and the reference gate increased significantly after motion correction, indicating improved image similarity. For the group study, NCC increased from an average of 0.522 to 0.681.
- Enhanced Image Quality: Qualitative results (Fig. 2) show that the motion-corrected images exhibit considerably reduced motion blur while maintaining acceptable noise levels, combining the benefits of gating (minimal blur, high noise) and no correction (acceptable noise, motion blur).
- Computational Time: The entire motion correction process for one patient took approximately 8 minutes on a standard multi-core PC.
- Diffeomorphic Transformations: The Jacobian determinants of the transformations were within acceptable ranges (0.81-1.19 for respiratory, 0.55-1.31 for cardiac), confirming that the transformations were diffeomorphic.
Limitations and Future Work
- Assumption of Independence: The paper's underlying assumption that cardiac and respiratory motion are independent is stated to be "approximately fulfilled for our data." While this simplification works well, it might not hold true in all cases, and further investigation into potential correlations could be beneficial.
- Further Optimization: The authors suggest that the registration could potentially be improved by performing "outer loops" of respiratory and cardiac motion correction, allowing for iterative refinement.
- Clinical Applicability: The authors believe their simplified pipeline is a significant step towards making motion correction methods more clinically applicable, especially as finer subdivision schemes become feasible due to the computational savings.
Summary updated