Enhancing and Denoising Mammographic Images for Tumor Detection using Bivariate Shrinkage and Modified Morphological Transforms

Abstract

Breast cancer stands as a prevalent concern for women worldwide. Mammography serves as the frontline defense for early detection, yet its low X-ray dosage often leads to suboptimal image quality. This study proposes a multi-step solution: (i) Image enhancement employs a two-step approach: denoising using bivariate shrinkage and a hybrid median filter based on stationary wavelet transform (SWT) to avoid shift variants, and it is combined with modified morphology operations including the background, a vignette image with the weighting function 1/R2. (ii) Segmentation utilizes the fast K-means algorithm with a straightforward technique to select the number of clusters and tumors automatically, within the segment containing the largest centroid. (iii) Classification employs a boosting ensemble model, based on statistical features extracted from SWT coefficients at different levels, for tumor classification to achieve reliable results. Utilizing mammograms from the MIAS (Mammographic Image Analysis Society) public dataset, the proposed method was tested on Gaussian noisy images, demonstrating superior performance compared to existing algorithms in detecting lesions. The segmentation achieves a high accuracy, 98.15% on average and a specificity of 99.56%. However, the ground truth occasionally extends beyond the tumor mass, resulting in a low sensitivity of 62.81%. Finally, classification is also performed using boosting ensemble learning with accuracy of 100% for training, testing and real data.