A wavelet-based spectral method for estimating the (directional) Hurst parameter in isotropic and anisotropic non-stationary fractional Gaussian fields is proposed. The method can be applied to self-similar images and, in general, to d-dimensional data which scale. In the application part, the problems of denoising 2D fractional Brownian fields and classification of digital mammograms to benign and malignant are considered. In the first application, a Bayesian inference calibrated by information from the wavelet-spectral domain is used to separate the signal from the noise. In the second application, digital mammograms are classified into benign and malignant based on the directional Hurst exponents which prove to be discriminatory summaries.