Comparação de famílias wavelets para denoising de imagens mamográficas digitais

Abstract

There is a big interest of the scientific community with health organizations, to improve the diagnosis of breast cancer, since it is one of the main causes of death in adult women. The most used technique for this, especially for initial stages, is the mammographyc image analysis of the patient, so improving the quality of these images is a fundamental task and directly linked to the correct diagnosis. For this, mathematical processing tools are used to reduce the amount of noise present in the final image. In this paper we discuss the use of several wavelet families as a denoising technique and compare the best results for each type of breast pattern, separated according to the classification proposed by BIRADS ™. A database containing forty images, ten of each pattern, was processed in MatLab in order to obtain signal-to-noise ratio (SNR) quantification as well as peak signal-to-noise ratio (PSNR). These values of SNR and PSNR were used as a tool to evaluate the best performances. In the literature, the higher the SNR and PSNR values, the better the signal-to-noise ratio and, therefore, the less noisy is the final image. Initially, 26 tests were performed on four images (one of each breast pattern) with different wavelet families and levels. With the result of these tests it was noticed that, regardless of breast pattern, the worst and best results were common to all four images. Thus, eight wavelets were selected for the processing of the other images. The results of the forty images were analyzed and the best performance was obtained by the coiflet 3 wavelet at level 1: higher SNR in 82.5% of the images and higher PSNR in 92.5%. It was not related to any specific breast pattern.