Ultrasound imaging is a vital, non-invasive diagnostic tool, but it suffers from a fundamental physical limitation: speckle noise. Unlike the additive electronic noise found in digital cameras, speckle is an inherent, interference-based artifact.
Understanding why speckle occurs—and how modern computational techniques mitigate it—requires looking closely at the intersection of acoustic physics, statistical signal processing, and deep learning.
The Physics of Speckle Noise
As explained by Greg G. Slabaugh et al. in Information-Theoretic Feature Detection in Ultrasound Images, speckle arises when acoustic waves scatter off microscopic tissue structures that are smaller than the ultrasound pulse wavelength. These scattered waves interfere both constructively and destructively before returning to the transducer, producing a grainy, high-contrast texture.
Because this noise is multiplicative rather than additive, its variance scales directly with the underlying signal intensity. This characteristic makes standard denoising filters ineffective. As a result, speckle severely degrades image resolution, blurs anatomical boundaries, and masks subtle lesions, as noted by Heng-Da Cheng et al. in Automated breast cancer detection and classification using ultrasound images: A survey. For automated computer vision systems, this granularity creates spurious edges and signal dropouts, complicating downstream clinical tasks like organ segmentation and tumor classification (Saman Nouranian et al. in An Automatic Multi-atlas Segmentation of the Prostate in Transrectal Ultrasound Images Using Pairwise Atlas Shape Similarity; Caifeng Shan et al. in Guest Editorial: Deep Learning in Ultrasound Imaging).
Acoustic Pulse ──> Microscopic Tissue Structures (Sub-resolution)
│
├──> Constructive Interference ──> Bright Pixels
└──> Destructive Interference ──> Dark Pixels
│
└──> Granular Speckle Pattern
Edge-Preserving Spatial Filters
To remove speckle without destroying crucial clinical details like tissue boundaries, researchers initially turned to advanced statistical filtering. Standard linear smoothing (such as Gaussian blurring) fails because it treats noise and edges identically, washing out the sharp transitions needed for diagnosis.
Anisotropic diffusion solves this by making the smoothing process conditional on local image gradients. The filter smooths heavily in homogenous regions but stops at boundaries where gradients are high. Shen Liu et al. demonstrated this in An anisotropic diffusion filter based on multidirectional separability by combining Speckle Reducing Anisotropic Diffusion (SRAD) with multidirectional separability to suppress heavy speckle while preserving sharp tongue contours. To make these filters less sensitive to over-iteration, Xiaoming Liu et al. introduced a robust Tukey error norm in A robust detail preserving anisotropic diffusion for speckle reduction in ultrasound images, ensuring that true edges act as absolute barriers to diffusion.
Non-Local Similarity and Low-Rank Recovery
A parallel intuition relies on non-local self-similarity: the observation that a patch of tissue often looks like other patches distributed elsewhere in the image. Instead of filtering purely within a local neighborhood, non-local methods search wide areas for structural matches.
[ Local Patch ] ── Searches Image ──> [ Matching Patch 1 ] [ Matching Patch 2 ]
│
Stacked into a Matrix
│
┌─────────────────────┐
│ Clean Structure │ ──> Low-Rank Component
│ (Highly Correlated) │
├─────────────────────┤
│ Random Speckle │ ──> High-Rank Noise
│ (Uncorrelated) │
└─────────────────────┘
By grouping these similar patches into a single matrix, the underlying clean anatomy forms a highly correlated, low-rank matrix, while the random speckle noise populates the higher ranks. In A Non-local Low-Rank Framework for Ultrasound Speckle Reduction, Chi-Wing Fu et al. leveraged this matrix structure, using a truncated weighted nuclear norm to separate the low-rank anatomical features from high-rank noise. Similarly, Shujun Liang et al. utilized a symmetric Kullback-Leibler (KL) divergence within a non-local total variation framework in Nonlocal total variation based on symmetric Kullback-Leibler divergence for the ultrasound image despeckling to measure patch similarity statistically, effectively distinguishing true tissue textures from random acoustic artifacts.
Multi-Scale and Frequency-Domain Decomposition
Because speckle manifests across multiple spatial scales, analyzing ultrasound images purely in the spatial domain can be limiting. Decomposing an image into different frequency sub-bands allows algorithms to target denoising where it is most effective.
The underlying intuition is that low-frequency bands capture broad anatomical structures, while high-frequency bands contain fine details mixed with heavy speckle. In Frequency division denoising algorithm based on VIF adaptive 2D-VMD ultrasound image, Hongbo Yan et al. applied an adaptive two-dimensional variational mode decomposition (2D-VMD) to split the image into distinct bands. This allowed the authors to apply anisotropic diffusion to low frequencies and BM3D filtering to high frequencies, balancing noise suppression and detail preservation.
In modern neural networks, this multi-scale intuition is embedded directly into network layers. For example, S.C. Liu et al. integrated wavelet convolutions into a U-Net architecture in Prediction Multiscale Cross-Level Fusion U-Net with Combined Wavelet Convolutions for Thyroid Nodule Segmentation, ensuring that fine-grained nodule boundaries are preserved during downsampling rather than being obliterated by noise.
Deep Learning for Noise Suppression and Delineation
With the advent of deep learning, the focus has shifted from explicit mathematical modeling of speckle to designing neural network architectures that inherently resist or suppress it. Standard convolutional networks often struggle with ultrasound because speckle noise can be mistaken for true semantic features.
To combat this, modern architectures incorporate specialized noise-suppression and boundary-refinement modules:
- Explicit Decoupling: In NSBR-Net: A Novel Noise Suppression and Boundary Refinement Network for Breast Tumor Segmentation in Ultrasound Images, Y. F. Sun et al. introduced a network that explicitly decouples noise suppression from boundary delineation, preventing blurred segmentations.
- Temporal Consistency: For dynamic applications like echocardiography, Huisi Wu et al. utilized region proxies and learnable kernels in Super-efficient Echocardiography Video Segmentation via Proxy- and Kernel-Based Semi-supervised Learning to maintain temporal consistency across video frames, preventing speckle from causing erratic frame-to-frame segmentation jitter.
- Manifold Learning: Unsupervised representation learning has also proven highly effective. Bo Li et al. demonstrated in Denoising convolutional autoencoder based B-mode ultrasound tongue image feature extraction that a Denoising Convolutional Autoencoder (DCAE) can be trained to reconstruct clean morphological features from corrupted images, forcing the latent space to encode robust anatomical structures rather than transient noise.
By embedding speckle-robust distance metrics directly into optimization frameworks—such as the graph-based patch representation and continuous min-cut approach by Anca Ciurte et al. in Semi-Supervised Segmentation of Ultrasound Images Based on Patch Representation and Continuous Min Cut—modern deep learning models can reliably segment highly degraded volumes, such as fetal femurs (Xu Wang et al. in Joint Segmentation and Landmark Localization of Fetal Femur in Ultrasound Volumes) or thyroid glands (Yishan Hu et al. in Ultrasound thyroid segmentation based on segmented frequency domain and local attention), matching or exceeding human expert performance.
Key Foundational Work
Automated breast cancer detection and classification using ultrasound images: A survey (2009) — contributed 1 concept elements to this direction.
Information-Theoretic Feature Detection in Ultrasound Images (2006) — contributed 1 concept elements to this direction.