Enhanced Wavelet Block Shrinkage Technique For Mammogram Denoising Using K-Means Clustering And Neural Networks

Abstract

The proposed research study introduces a novel approach for denoising digital mammograms by improving the existing wavelet block shrinkage filtering method with K-Means clustering and a convolutional neural network. This approach involves decomposing both the original and noisy mammograms into frequency subbands using 2D discrete wavelet transformation. The resulting subbands are then grouped into multiple clusters based on similar features of the wavelet coefficients, employing K-Means clustering. This represents an improvement over the traditional block shrinkage method, which uses fixed-size blocks. These clusters from the original and noisy mammograms are paired to train a convolutional neural network, which serves as an optimal shrinkage function. This neural network-based thresholding mechanism replaces traditional hard and soft thresholding methods that rely on a universal threshold. Test results demonstrate that the proposed enhanced wavelet block shrinkage mechanism achieves a 20% improvement in peak signal-to-noise ratio and a 5% increase in structural similarity index score compared to traditional wavelet block shrinkage.