Abstract
Ultrasound imaging of the tongue is a critical tool in speech science, yet its utility is often limited by pervasive speckle noise. This paper introduces a novel anisotropic diffusion filter that enhances image clarity by combining Speckle Reducing Anisotropic Diffusion (SRAD) with a multidirectional statistical separability metric. Unlike traditional gradient-based methods, this approach employs Linear Discriminant Analysis to evaluate the statistical distinctness of neighboring regions, providing a more robust edge-detection mechanism in high-noise environments. Evaluations on synthetic and real-world ultrasound data show that the proposed filter significantly outperforms standard SRAD in noise reduction and edge preservation accuracy.
Introduction
In speech research, B-mode ultrasound provides a non-invasive, real-time window into the complex movements of the tongue. However, the resulting images are typically degraded by speckle noise—a multiplicative interference pattern caused by the scattering of sound waves. This noise obscures the sharp contours of the tongue, making both manual tracing and automated analysis difficult. While Speckle Reducing Anisotropic Diffusion (SRAD) is a common solution, its reliance on local intensity gradients often leads to "bumpy" results or blurred edges when noise levels are high.
Methodology
The core contribution of this work is the integration of multidirectional separability into the partial differential equation (PDE) governing image diffusion.
Statistical Separability
Traditional filters use gradients (the difference in brightness between adjacent pixels) to identify edges. In contrast, this method uses statistical separability. By comparing the distribution of pixel intensities in two adjacent regions (Region 1 and Region 2), the filter can determine if they belong to the same tissue or are separated by a boundary.
This approach, rooted in Linear Discriminant Analysis, is far more resilient to the grainy nature of speckle noise than simple gradient calculations.
Multidirectional Framework
To further refine edge detection, the authors expanded the 1D separability concept into a 2D multidirectional framework. The filter calculates separability across four distinct regions surrounding a central pixel $(i, j)$.
By weighting these directional components based on the local gradient direction, the filter becomes highly sensitive to the specific orientation of the tongue's surface. The diffusion process is dynamically adjusted: smoothing is aggressive in homogeneous areas but becomes "sluggish" (stops) when multidirectional separability is high, thereby preserving critical anatomical boundaries.
Experimental Evaluation
The proposed filter was evaluated using two datasets:
- Synthetic Data: 400×240 pixel images with predefined gray levels and simulated speckle noise. This allowed for objective benchmarking against a known "ground truth."
- Real Ultrasound Data: Captured using a Terason T3000 machine. Images were cropped to 310×190 pixels to focus on the tongue anatomy and optimize processing speed.
Results
The proposed method demonstrated clear quantitative and qualitative advantages over standard SRAD:
- Denoising Efficiency: On synthetic images, the Mean Squared Error (MSE) was reduced from 49.96 (SRAD) to 42.15. Peak Signal to Noise Ratio (PSNR) showed a corresponding improvement.
- Edge Preservation: The Figure of Merit (FOM), which measures how closely a detected edge matches the ideal boundary, rose from 0.8123 for SRAD to 0.9144 for the proposed method.
- Contour Extraction: When applied to real ultrasound frames, the filter enabled automated tools like Active Appearance Models (AAM) to trace tongue contours with significantly higher completeness and accuracy.
Conclusion
By shifting the focus from simple pixel gradients to regional statistical distributions, the multidirectional separability filter provides a robust solution for ultrasound denoising. While the method introduces additional computational complexity, its ability to preserve sharp edges in the presence of heavy speckle noise makes it an essential preprocessing step for automated tongue tracking and speech analysis systems.
Summary updated