Multi-Resolution Wavelet Packet-Driven Dual Path CNN for Breast Lesion Classification

Breast cancer diagnosis, based on histopathological imaging, is essential for clinical decision-making due to complex nature of breast cancer subtypes, diverse morphological characteristics of lesions, and critical need for precise classification. However, accurate lesion classification remains challenging due to image noise, high intra-class variability, and inter-class similarity among different lesion types. Current methods often lack robustness when dealing with multi-resolution images, varying cell structures, inconsistent staining, and overlapping tissue regions. To address these challenges, we propose a Multi-resolution Wavelet Packet Transformation with Dual-path Convolutional Neural Network (DWT-DPCNN) for accurate classification of breast lesions in histopathological images. In the initial stage, after standard pre-processing steps (resizing and rescaling), the images undergo a two-level 2D discrete wavelet transformation. Subsequently, two distinct convolutional neural network blocks are employed: a residual identity block-based CNN, which effectively preserves crucial information, and a dense block-based CNN, which facilitates efficient feature propagation. The resulting low-level sub-band images, generated using a Haar wavelet filter, are fed into the DWT-DPCNN model to enable precise differentiation between benign and malignant cells in the histopathological images. Extensive experiments demonstrate the efficacy of our framework, achieving an accuracy of up to 96.02±1.72% in the classification of breast lesions. Further experiments conducted utilizing four distinct image resolutions of 40×, 100×, 200×, and 400×, indicate that the proposed method attains the highest performance in both lesion classification and diagnosis, with an accuracy of up to 96.54±1.01%, underscoring its promising potential for clinical applications.