TL;DR
This paper introduces two automated methods, Radon transform and wavelet analysis, for tracking muscle fascicle orientation in ultrasound images. These methods offer a more objective and efficient alternative to manual digitization, crucial for understanding muscle function during dynamic movements.
What Problem Does This Paper Address?
Understanding muscle architecture, particularly the orientation of muscle fascicles, is vital for comprehending how muscles generate force and move the body. Traditionally, this has been done by manually tracing fascicles in ultrasound images, a process that is both time-consuming and prone to subjective errors. Fascicle orientation can also change dynamically during muscle contractions, and these changes are difficult to capture accurately with manual methods. The paper aims to develop automated techniques to overcome these limitations, enabling more precise and efficient analysis of muscle fascicle orientation, especially during dynamic activities.
Methods
The study employs a two-stage approach:
- Multiscale Vessel Enhancement Filtering: This initial step enhances the visibility of fascicle structures, which appear as dark lines, against the brighter connective tissue. It uses a series of Gaussian kernels to effectively highlight tubular structures of varying sizes and even curved ones, while simultaneously reducing image noise.
- Fascicle Orientation Quantification: Two distinct methods are then applied to the filtered images:
- Radon Transform: This method is used to identify the dominant orientation of fascicles within an entire image. It works by projecting parallel lines across the image and calculating intensity integrals along these lines, with variations in these integrals indicating the dominant orientation.
- Anisotropic Wavelet Analysis: This technique quantifies the local orientation of fascicles at each pixel. It utilizes a specially designed wavelet kernel that is sensitive to specific orientations, allowing for a more detailed, pixel-by-pixel analysis of fascicle direction.
Both methods are validated against synthetic images with known orientations and real ultrasound data that was manually digitized by multiple researchers.
Datasets
The study utilizes two types of datasets:
- Synthetic Images: These are generated grids of parallel lines with controlled orientations and varying levels of added noise. They serve as a benchmark to assess the accuracy of the automated methods under controlled conditions.
- Real Ultrasound Images: B-mode ultrasound images of the vastus lateralis muscle were recorded from a subject during cycling. These images were manually digitized by ten researchers to provide a ground truth for validation.
Key Insights and Ideas
- Automated vs. Manual is a Trade-off: While automated methods offer objectivity and speed, the inherent variability in manual digitization (as shown by the 1.41° standard deviation among researchers) means that direct comparison needs careful interpretation. The automated methods might not perfectly match manual results simply because the manual results themselves are not perfectly consistent.
- Radon for Dominant, Wavelet for Local: The Radon transform is excellent for capturing the overall, dominant fascicle direction, which is useful for tasks like estimating fascicle length. The wavelet analysis, on the other hand, provides a more granular, pixel-level understanding of orientation, which is crucial for analyzing fascicle curvature and regional variations.
- Noise Sensitivity: Both methods show increased error with higher noise levels in synthetic images, but they remain robust for contrast-to-noise ratios typical of real ultrasound data.
Key Results
- High Accuracy on Synthetic Data: Both Radon and wavelet methods achieved very low mean absolute errors (0.058° and 0.02°, respectively) on synthetic images across a range of orientations, indicating high precision.
- Robustness to Noise: Even with increased noise, errors remained below 0.02° for contrast-to-noise ratios greater than 0.8, a level comparable to or better than what is observed in real ultrasound images.
- Wavelet vs. Radon on Real Data: On real ultrasound data, the Radon transform showed a closer match to manual digitization results than the wavelet analysis. However, the wavelet analysis provided more detailed information about local orientations and curvature.
- Manual Digitization Variability: The standard deviation of fascicle orientations from manual digitization by ten researchers was 1.41°, highlighting the subjective nature of manual methods.
Limitations and Future Work
- Manual Digitization Accuracy: The paper acknowledges that the accuracy of manual digitization is not perfectly known, making it difficult to definitively state which automated method is "more accurate" when compared to manual results. Future work could involve developing more standardized manual digitization protocols or using a larger cohort of experts.
- Wavelet's Local Detail: While the wavelet analysis provides rich local orientation data, its direct comparison to manual methods for overall orientation estimation showed a statistically significant difference (-1.35°). Further investigation into why this difference occurs and how to best leverage the local information for global estimations is warranted.
- Broader Application: The study focused on the vastus lateralis. Future work could explore the applicability of these methods to other muscles and different imaging conditions. The authors suggest that the combined multiscale vessel enhancement and wavelet analysis methods offer a promising avenue for quantifying local fascicle orientations and non-linear trajectories, which could be further explored.
Summary updated