Log exponential shrinkage: a denoising technique for breast ultrasound images

The process of uncontrolled fatal growth of tissue that invades its surrounding parts is called cancer. The cancer in breast is the most commonly encountered cancer among women. It accounts for millions of deaths around the globe. Early detection of breast cancer by ultrasound (US)-based diagnosis can have major positive influence on life expectancy of a patient. Unlike other diagnosis methods, the US-based diagnosis is harmless, economical, and allows frequent use. Only limitation it has is the presence of noise. So, we propose an US denoising technique named log exponential shrinkage (LES). In LES, US image was initially processed for additive noise. A scale-based thresholding function and a novel speckle noise variance estimator are the main contribution of this study. The estimator utilizes sub-band coefficients of a transformed image to evaluate the noise intensity. The threshold to filter out the noisy coefficient was calculated using the noise variance. Ultimately, the thresholding exponential function was used to modify noisy coefficients. The efficiency of LES was compared with other denoising techniques in terms of no-reference, full reference quality metrics and edge preservative metrics. The performance was tested on one synthetic image and two breast cancer datasets. To comprehend the extent of advancement by LES, a percentage improvement table has been provided in the study, having 300% better results. The novel estimator has improved the error in estimation by 95.8% at the optimal noise intensity. With such promising results, LES might find application in breast cancer diagnosis using US.