Power Spectral Fractal Dimension and Wavelet Features for Mammogram Analysis: A Machine Learning Approach

The paper delineates a novel method based on power spectral fractal dimension for the identification, classification, and prediction of normal, benign, and malignant regions in a mammogram. For this, the fractal dimension values are found in the radial directions from a reference point and are used for classifying the regions, unlike conventional methods. Malignant lesions show a higher fractal dimension compared to benign and normal lesions. The study is extended to the classification of normal, benign, and malignant regions based on wavelet features using the machine learning techniques—cubic support vector machine and neural network pattern recognition. For the wavelet feature-based classification, a lossless image enhancement is realized through the contrast limited adaptive histogram equalization method, and the texture features are analyzed from the gray level co-occurrence matrix. Multiresolution analysis based on the wavelet transform method was used to enhance the high-frequency components of the mammograms. The Daubechies 8, level-4 wavelet coefficients, and the texture features serve as the input variables to the cubic support vector machine and neural network pattern recognition, which classify and predict the lesions with an accuracy of 80 and 96.7%, respectively, and suggest its potential in mammogram analysis.