TL;DR
Researchers have developed a semi-automatic pipeline for Diffusion Tensor Cardiovascular Magnetic Resonance (DT-CMR) that uses low-rank image decomposition to stabilize motion correction. By creating denoised, motion-stripped reference frames and implementing an automated "outlier" rejection system, the method significantly improves the quality of cardiac fiber orientation maps while reducing the need for tedious manual data cleaning.
What Problem Does This Paper Address?
Imaging the microscopic structure of a beating heart is a monumental challenge. DT-CMR allows us to visualize the orientation of cardiomyocytes (heart muscle cells) in vivo, which is vital for understanding heart disease. However, the process is plagued by two major issues:
- Motion Artifacts: The heart is constantly moving due to both the cardiac cycle and the patient's breathing. Even with breath-holds and cardiac gating, images often don't align perfectly.
- Contrast Variability & Noise: DT-CMR requires multiple "diffusion-encoded" images. These images look very different from one another and often have a low signal-to-noise ratio (SNR), making standard image registration (alignment) algorithms fail.
Traditionally, researchers have to manually crop images and hand-pick "good" frames, discarding those with signal loss or through-plane motion. This is subjective, slow, and prone to error.
Methods
The authors propose a two-stage pipeline to automate and robustify this workflow:
- Low-Rank Decomposition: They treat the stack of acquired images as a multidimensional tensor and decompose it using the Tucker form.
- The Reference: They use the geometric average of Rank-1 images as the registration reference. Because Rank-1 captures only the most consistent spatial information, it effectively "disentangles" the image from respiratory shifts and local deformations.
- The Moving Frames: They use Rank-6 denoised images as the "moving" frames for registration. This suppresses noise while retaining enough detail for accurate alignment.
- Automatic Frame Selection: Instead of manual review, the system defines a "donut-shaped" Region of Interest (ROI) around the myocardium. It then calculates the Pearson correlation between frames. Frames that fall more than three scaled median absolute deviations below the median are automatically rejected as outliers (likely caused by signal loss or severe misregistration).
- Registration: The actual alignment is performed using the
itk-elastixlibrary, applying affine and rigid transformations derived from the low-rank images back to the original data.
Datasets
The study evaluated the method using data from:
- 12 healthy volunteers scanned at both 1.5T and 3T field strengths.
- Two common DT-CMR sequences: STEAM (Stimulated Echo Acquisition Mode) and SE-EPI (Spin Echo Echo Planar Imaging).
- Data included $b_0$ images and 6 encoding directions at $b = 150$ and $600\text{ s/mm}^2$, with more than 9 averages per slice.
Key Insights and Ideas
- Rank-1 as a "Motion Filter": The most clever insight is using Rank-1 images for the reference. In a noisy, moving stack of images, the "Rank-1" component represents the underlying static structure. By registering everything to this "essence" of the heart, the algorithm isn't distracted by transient noise or breathing artifacts.
- Denoising over Cropping: Standard workflows often require aggressive manual cropping to remove high-intensity signals from adjacent organs (like the liver or lungs) that confuse registration. The low-rank denoising (Rank-6) makes the registration robust enough that manual circular cropping is no longer necessary.
- Generalization: While automatic frame rejection was originally designed for specific sequences (STEAM), this paper proves it generalizes well to SE-EPI data, matching the performance of human experts.
Key Results
The proposed "Low rank + auto" method was compared against a traditional "DFT + manual" (Discrete Fourier Transform registration + manual selection) workflow:
- Better Fit: The R-square values for the Helix Angle (HA)—a key metric for cardiac fiber orientation—were higher across the board (e.g., 0.925 for STEAM vs. 0.914 for the traditional method).
- Lower Error: The Root Mean Square Error (RMSE) was consistently lower in the proposed pipeline.
- Improved Physical Validity: A major "stress test" for DT-CMR is the number of negative eigenvalues in the diffusion tensor (which are physically impossible and indicate bad data). The proposed method significantly reduced these "per mile" occurrences (e.g., from 11.3‰ down to 5.10‰ in STEAM).
Limitations and Future Work
- Centering Issues: The automatic cropping still struggles if the heart is not centrally located in the initial field of view, sometimes requiring a quick manual adjustment.
- Computational Trade-off: While the method is more robust, it is slower than simple DFT-based registration. Processing a stack of 63 slices took about 20 seconds, compared to the near-instantaneous (0.2s) performance of simpler methods.
- Future Directions: The authors suggest that further refining the automatic segmentation and centering could lead to a truly "push-button" DT-CMR post-processing suite.
Summary updated