Wavelet-CNN Feature Fusion Architecture for Robust Breast Cancer Classification in Histopathological Imaging

Abstract Breast cancer remains a leading cause of cancer-related mortality among women, underscoring the critical need for early and accurate diagnostic strategies. In this study, we introduce a multiscale feature fusion framework that integrates the Lifting Wavelet Transform (LWT) with a Multi-Path Convolutional Neural Network (CNN) to enhance the detection of breast cancer in histopathological images. The proposed methodology is evaluated using the BreakHis dataset, which comprises 7,638 histopathology images captured at four magnification levels (40×, 100×, 200×, and 400×), each annotated as either benign or malignant. The processing pipeline begins with image normalization and resizing, followed by the application of a three-level two-dimensional LWT to extract informative low-frequency sub-band components. These wavelet-derived features are subsequently fed into a custom-designed multi-path CNN, where distinct convolutional branches are dedicated to processing features specific to each decomposition level, thereby facilitating more effective lesion classification.Comprehensive experimental analysis demonstrates that the proposed framework achieves superior diagnostic accuracy, outperforming established pre-trained CNN models. Notably, the network attains a testing accuracy of 99.28% when combining images at 40×, 200×, and 400× magnification levels using the Haar wavelet filter. These results substantiate the efficacy of multiscale wavelet-CNN feature fusion for histopathological breast cancer detection, offering a robust approach for early and reliable diagnosis.