IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 18, NO. 9, SEPTEMBER 1999 787 A Novel Multiscale Nonlinear Thresholding Method for Ultrasonic Speckle Suppressing Xiaohui Hao,* Shangkai Gao, Senior Member, IEEE, and Xiaorong Gao Abstract— This paper presents a novel speckle suppression method for medical B-scan ultrasonic images. An original image is first separated into two parts with an adaptive filter. These two parts are then transformed into a multiscale wavelet domain and the wavelet coefficients are processed by a soft threshold- ing method, which is a variation of Donoho’s soft thresholding method. The processed coefficients for each part are then trans- formed back into the space domain. Finally, the denoised image is obtained as the sum of the two processed parts. A computer- simulated image and an in vitro B-scan image of a pig heart have been used to test the performance of this new method. This technique effectively reduces the speckle noise, while preserving the resolvable details. It performs well in comparison to the multiscale thresholding technique without adaptive preprocessing and two other speckle-suppression methods. Index Terms— Multiscale thresholding, ultrasound speckle, wavelet transform. I. INTRODUCTION U LTRASONIC images suffer from a special kind of noise called speckle. Speckle is a term used for the granular pattern that appears in B-scan images and can be considered as a kind of multiplicative noise. It is caused by the constructive and destructive interference of back-scattered signals due to unresolved tissue inhomogeneity. It should be emphasized that speckle is a consequence of unresolved scatters, which are caused by tissue-ultrasound interaction, and not an image of the scatters. Speckle occurs especially in images of the liver and kidney whose underlying structures are too small to be resolved by large wavelength ultrasound. It significantly degrades the image quality and increases the difficulty in discriminating fine details in images during diagnostic exami- nations. It also complicates further image processing, such as image segmentation and edge detection. Ultrasound speckle-reduction techniques can be classified into two types: compounding and filtering [1]. In the com- pounding technique, a series of images of one target are sampled at different times, with different ultrasound frequen- cies, or in different scan directions. They are then merged to form a composite image. This method suffers from a decrease in space resolution. In filtering techniques, various filters are used to reduce speckle noise in ultrasonic images. One of them is the Wiener Manuscript received February 10, 1998; revised August 24, 1999. This work was supported in part by the Beijing Nature Science Foundation. The Associate Editor responsible for coordinating the review of this paper and recommending its publication was M. Insana. Asterisk indicates corresponding author. *X. Hao, S. Gao, and X. Gao are with the Department of Electrical Engineering, Tsinghua University, Beijing, 100084, China. Publisher Item Identifier S 0278-0062(99)09058-8. filter. Since it is developed mainly for additive random noise reduction, it has little success in speckle suppression. Local adaptive filters [2], [3] are also used. The weights of the adap- tive filter are determined from local statistics. This method can suppress speckle effectively. However, it fails to sufficiently preserve the weak and diffuse edges in the images. In the third method (in this paper, this method is called DCSS) an ultrasound image is transformed into the wavelet domain and components in some scales are discarded. After the inverse wavelet transform, the denoised image is acquired. In DCSS, if only low-scale details are discarded, the speckle cannot be successfully suppressed and if details of the high scale are discarded, useful signals may be lost. (In the wavelet domain, the high scale corresponds to the low frequency.) Donoho presented a soft-thresholding denoising method in 1995 [4]. In this method, the observed signal is decomposed into the wavelet domain. The coefficients are then processed by thresholding. Finally, the estimated image is reconstructed by taking the inverse wavelet transform of the processed coefficients. Xiang and Zhang utilized this method to reduce ultrasound speckle [5]. They used the same threshold in all scales to maximize the SNR, but this may not match the specific distribution of signal and noise in different scales. Zong et al. [8] presented a multiscale homomorphic ap- proach for speckle suppressing and feature enhancement. Un- der the multiplicative speckle model, they used a logarithmic transform to change signal-dependant speckle to additive white noise. Then regularized wavelet soft thresholding (wavelet shrinkage) was used to remove noise energy in the finer scales and nonlinear processing of feature energy was used to enhance image contrast. All of the above methods failed to balance between the speckle suppression and detail signal preservation. In this paper, we propose a novel multiscale nonlinear thresholding method with adaptive preprocessing (MSNLT-A) to suppress speckle (in Section II). For comparison, speckle suppression by multiscale thresholding without preprocessing (MSNLT) is also described. Experimental results of the MSNLT-A are compared with those of MSNLT, adaptive filtering, and DCSS techniques (in Section 3). II. THE MSNLT-A METHOD A. Introduction As mentioned above, the adaptive weighted median filter (AWMF) developed by Loupas, [1] and Karaman et al. [3] effectively suppresses speckle. However, it loses many useful details because it is merely a low-pass filter. When being used 0278–0062/99$10.00  1999 IEEE