Spatial Transformer Network with Deep Features for Improved Breast Cancer Image Classification

Breast cancer continues to be an utmost worldwide health issue, accounting for a notable number of cancer-related deaths among women every year. premature and precise identification is vital to improve prognosis and reduce mortality. However, automated classification of mammographic images is challenged by low-resolution data, spatial distortions, and the lack of fine-grained texture characteristics critical for identifying malignant tissues. In this research, we present a revolutionary deep learning framework that conveys these limitations through an integrated architecture combining a Super-Resolution Generative Adversarial Network (SRGAN), a Spatial Transformer Network (STN), and a ResNet50-based image classifier. A significant advancement in our architecture is the incorporation of embedding Bilinear Interpolation (BI) and Lanczos interpolation (LZ) within the STN module. Unlike traditional fixed-kernel Lanczos filters, our learnable version enables the network to adapt interpolation weights during training, allowing it to better handle spatial variations and preserve critical anatomical details. The SRGAN enhances the resolution of mammographic images, aiding in improved feature extraction, while the ResNet50 backbone performs classification with refined spatial awareness. The enhanced images are then passed through the STN for spatial normalization, followed by classification using a fine-tuned ResNet50 backbone. We evaluated the proposed pipeline on the MIAS (Mammographic Image Analysis Society) and Digital Database for Screening Mammography (DDSM) dataset, a benchmark in breast cancer imaging research. The model demonstrated superior classification performance, achieving 97.08% accuracy, 95.91% precision, 95.08% specificity, and 95.48% f-measure on DDSM dataset and 96.58% accuracy, 94.62% specificity, 94.59% sensitivity, and an f-measure of 94.59% on MIAS datatset. These outcomes show a significant improvement over comparable approaches that employ traditional interpolation or omit spatial correction mechanisms.