A Neural Network-Based Wavelet Thresholding Technique for De-Speckling of Ultrasound Images

The mortality rate and life lost in suffering can be significantly improved if the cancer is detected in its earliest stage. For early detection, it is required to design a system that can be used frequently. Only ultrasound waves that are used for disease diagnosis are free from radiation. The ultrasound images suffer from speckle noise, which reduces their visual quality. Machine learning techniques have helped researchers to achieve significant results. Its application has been extended to almost every field of science. So, reducing the speckle noise in ultrasound images by using machine learning techniques is the aim of this study. In the proposed methodology, the image is transformed using logarithmic and wavelet transform. The histogram of the detailed sub-band was then used to train a neural network to generate modified coefficients. The network was trained in an unsupervised fashion. The study was tested on the breast and liver ultrasound images. The de-speckling of computer-generated images resulted in a PSNR and MSE of 45.6 and 58, respectively. The technique when applied to the ultrasound image dataset the reference less quality evaluator gave the value 31.6, which is comparatively better than the original image.