About Medical Ultrasound Images (Part 2)

In the imaging process of a medical ultrasound imaging system, the ultrasonic waves generated by the ultrasound transducer are reflected back to the transducer after passing through the human body interface. The scattering of rough surfaces of organs and randomly distributed scatterers within soft tissues will form a series of coherent waves, causing random fluctuations in the signals on the transducer, forming speckle noise. This is the main source of noise in ultrasound images. The speckle noise superimposed on the tissues prevents the ultrasound image from correctly reflecting the characteristics of the target, leading to reduced image contrast and seriously affecting the difficulty for clinicians to distinguish between normal and diseased tissues. It also poses challenges for subsequent processing such as edge detection and image segmentation.

Speckle noise is inherent in all coherent imaging systems. When ultrasound waves encounter the interface of human tissue structures, due to acoustic impedance mismatch, some energy will be reflected. If the size of the human tissue structure is close to or smaller than the wavelength of the incident ultrasound wave, ultrasound scattering occurs. The scattered echoes with different phases interfere with each other, producing speckle noise. Larger objects, however, produce coherent reflections, and these reflected signals reflect the information of the human tissue structure.

Precisely establishing a speckle noise model can better describe the characteristics of the image. There are usually two types of medical ultrasound images: one is obtained directly from the received RF signal, and the other is a display ultrasound image after dynamic range compression. Due to the non-linear nature of dynamic range compression, the statistical characteristics of these two types of images are different.

The dynamic range of ultrasound echo signals is very large, so medical ultrasound images usually undergo dynamic range compression to meet the needs of actual image display and physician visual observation. However, dynamic range compression alters the statistical characteristics of the ultrasound echo signals.

Conclusions from the speckle noise model: For medical ultrasound images that have not undergone dynamic range compression, the brightness distribution of their speckle noise is asymmetrical; for medical ultrasound images that have undergone dynamic range compression, the brightness distribution of their speckle noise is approximately symmetrical, but it may also be asymmetrical. Therefore, in medical ultrasound image noise suppression, non-linear denoising methods perform better.

The main difficulties in ultrasound image denoising are:

1)Speckle noise can be roughly regarded as a multiplicative noise.；

 2)The stochastic nature of the noise is quite complex;

3)Noise is easily confused with image details, and image details are complex and diverse.

Existing speckle noise suppression algorithms are roughly divided into: median filtering; spatial domain local statistical filtering algorithms; anisotropic diffusion filtering algorithms; and multi-scale transform-based filtering algorithms (wavelet-based filtering algorithms).

Median filtering methods automatically select the point weights, window size, and shape within the filtering window based on the local statistical characteristics of the image. Although they have achieved certain results in preserving image details, they are very sensitive to window selection, which limits the processing effect.

Wavelet transform-based methods transform the image into the wavelet domain, use wavelet threshold processing to discard coefficients considered to be noise, and then inversely transform back to the image. However, there is no definitive method for selecting the scale and threshold of the wavelet transform.

The anisotropic diffusion algorithm is based on partial differential equations. It introduces the theory of partial differential equations into image processing. It has a larger diffusion coefficient at non-boundary points, resulting in more image smoothing, while at boundary points, the diffusion coefficient is smaller. The denoising effect largely depends on the accurate judgment of the image boundary. In addition, selecting a suitable diffusion cutoff time is a challenge for the anisotropic diffusion algorithm. If the original image is subjected to unlimited diffusion, the final image obtained will be a monochromatic image with the average gray value of the initial image.

This consultation comes from the internet. Please contact us for deletion if there is any infringement.