Content-Aware Seam-Carving Resizing With U-Net Segmentation: A Preprocessing Framework That Improves CNN-Based Breast-Cancer Detection in Mammograms

Mammography is widely used for the early detection of breast cancer. Although convolutional neural networks (CNNs) have achieved promising results in mammogram classification, preserving critical image features during down sampling for accurate segmentation of regions of interest (ROIs) remains an open challenge. In this study, we introduce a hybrid framework that integrates the Seam Carving Algorithm (SCA) for content-aware image resizing with U-Net-based segmentation as a preprocessing pipeline prior to CNN classification. The proposed approach first reduces the dimensions of mammographic images while preserving high-information structures and subsequently applies a U-Net model to segment breast tissue and lesions. These segmented outputs are then fed into downstream CNN-based classifiers. We evaluated the performance of three deep learning architectures—DenseNet121, ResNet50, and VGG16—on both film-based and digital mammography datasets under two distinct preprocessing strategies: i) Seam Carving followed by Bicubic Interpolation (BSCBI): High-resolution images (exceeding 2500 x 2000 pixels) were first downsized to 750 x 500 pixels using SCA and then resized to 256 x 256 pixels via bicubic interpolation and, ii) Single Bicubic Interpolation (BI): Images were directly resized from their native high resolution to 256 x 256 pixels using bicubic interpolation alone. The results demonstrate that the BSCBI preprocessing strategy consistently outperforms the standard BI approach across all evaluated models and imaging modalities, yielding classification accuracy and sensitivity scores above 98%, compared to approximately 92% for BI. Ultimately, combining Seam Carving-based downsizing with U-Net segmentation markedly improves the detection of malignant cases across all benchmark CNN models. This hybrid framework shows strong potential for advancing computer-aided and fully automated breast cancer screening systems.