Methodological approach to detecting pathological changes in mammograms using wavelet analysis

Abstract

Early detection of pathological changes in the breast is an important task in medical diagnosis, as pathological structures may have low local contrast, small size, and may be obscured by the heterogeneous anatomical structure of breast tissue at the early stages of disease. This highlights the relevance of developing methods for automated analysis of digital mammographic images capable of accounting for the spatial and multiscale characteristics of pathological regions. This study proposes a methodological approach to the automated detection and localization of pathological changes in digital mammograms based on the two-dimensional discrete wavelet transform. A reproducible algorithm with fixed parameters and an original suspiciousness map construction rule was developed. The algorithm includes breast region segmentation, Perona–Malik anisotropic diffusion, local contrast enhancement using the Contrast Limited Adaptive Histogram Equalization method, three-level wavelet decomposition, soft coefficient thresholding using the universal threshold, and calculation of local statistical and textural characteristics. The scientific novelty lies in the proposed inter-scale consistency criterion, according to which local suspiciousness is determined not only by the magnitude of an individual wavelet coefficient but also by the consistency of energy, statistical, and spatial-textural characteristics across adjacent scales. Separate rules are defined for different types of mammographic manifestations: microcalcifications are analyzed primarily using high-frequency energy and kurtosis; masses are characterized by low-frequency structure and texture; and architectural distortions are analyzed based on inter-directional and inter-scale heterogeneity. Local analysis windows, fixed weights of the suspiciousness map components, and post-processing rules are specified. The developed methodology has a modular structure, enabling individual components to be independently tuned and further improved. It provides a basis for implementing an automated mammogram analysis system whose output can consist of a localized set of candidate regions with their corresponding characteristics. Further research involves implementing the algorithm on publicly available annotated mammogram datasets, separately verifying three types of pathological manifestations, and analyzing sensitivity to the wavelet type and decomposition level. A practical outcome of further development of the methodology is the creation of a software module for automated mammogram analysis. Such a module can perform image preprocessing, multiscale wavelet decomposition, automatic generation of suspiciousness maps, and identification of candidate regions. Its output is not intended to establish a diagnosis but rather to provide a structured set of localized regions and their characteristics as additional information for subsequent image interpretation by a specialist. Further development of the methodology may focus on machine learning and deep learning. Wavelet-based features can be used as an input feature space for classical classification algorithms, while the generated suspiciousness maps can serve as a mechanism for preliminary region-of-interest identification for convolutional neural networks or other deep learning models. Another promising direction is the investigation of hybrid architectures that combine multiscale wavelet representations with automatically learned features. In such a framework, the proposed methodology can serve as an interpretable and parameterized preprocessing stage, complementing the capabilities of modern machine learning models.