Abstract
Accurate segmentation of ultrasound images is often hindered by speckle noise and intensity inhomogeneity. This paper introduces a novel active contour model that replaces standard Gaussian noise assumptions with a Local Rayleigh Distribution Fitting (LRDF) energy term. By tailoring the mathematical model to the specific statistical properties of ultrasound speckle, the proposed method achieves high precision in clinical applications, specifically in uterine fibroid segmentation, yielding a Dice Score Coefficient (DSC) of 0.95.
Introduction
Ultrasound imaging is a cornerstone of clinical diagnostics due to its non-invasive and real-time nature. However, automated segmentation remains challenging because of two primary factors: intensity inhomogeneity (uneven brightness) and speckle noise. Traditional active contour models typically assume that image noise follows a Gaussian distribution. While effective for modalities like MRI, this assumption fails for ultrasound because the physics of acoustic reflection produces speckle patterns that are more accurately described by a Rayleigh distribution. When models utilize incorrect statistical assumptions, they often "leak" across weak boundaries or fail to capture the complex morphology of organs.
Methodology
The core innovation of this research is the Local Rayleigh Distribution Fitting (LRDF) energy term. Unlike global models that analyze the entire image domain, the LRDF model calculates energy within local neighborhoods ($N_x$), allowing the contour to adapt to local contrast variations and intensity inhomogeneity.
Rayleigh Statistical Modeling
The model utilizes the Rayleigh distribution to characterize local pixel intensities. As shown in the statistical analysis of clinical regions of interest (ROI), the Rayleigh distribution provides a significantly better fit for ultrasound speckle than the standard Gaussian distribution.
Energy Minimization and Level Sets
The boundary is represented implicitly by a level set function ($\phi$). The total energy function minimized via gradient descent includes:
- The LRDF Term: The primary driver for boundary detection based on Rayleigh statistics.
- Penalty Term: Maintains the smoothness of the level set function, eliminating the need for computationally expensive re-initialization.
- Length Penalty Term: Ensures the resulting contour remains smooth and biologically plausible.
Experimental Results
The researchers evaluated the LRDF model using both synthetic B-scan simulations and in-vivo clinical images from eight patients with uterine fibroids. The results were benchmarked against the popular Local Binary Fitting (LBF) method and standard level set models.
The LRDF method achieved a Dice Score Coefficient (DSC) of $0.95 \pm 0.02$, indicating a near-perfect match with manual expert segmentations. In head-to-head comparisons, the LRDF model consistently outperformed the LBF model, which frequently over-segmented regions or missed boundaries in areas of low contrast.
Discussion and Conclusion
The success of the LRDF model highlights the importance of physics-informed mathematical modeling in medical imaging. By switching to a Rayleigh distribution and integrating energy locally, the model effectively handles the unique noise and lighting characteristics of ultrasound.
While the results are promising, future work should focus on expanding the dataset beyond the initial eight patients to ensure clinical generalizability. Additionally, further optimization of the double integral energy function could reduce computational overhead, making the method more suitable for real-time clinical applications across different organ systems, such as the liver or heart.
Summary updated