Improving Breast Cancer Detection Accuracy using Convolutional Block Attention Modules

This study's main objective focuses on enhancing the precision of breast cancer identification. There are two primary stages in this study's framework. First stage: compare the performance of two different convolution neural networks (CNNs) architectures with data splitting size representing 20% of the testing and validation set and 80% of the training data. The convolutional neural networks used in this study are VGG-19 (Visual Geometry Group 19) and ResNet50 (Residual Network 50), both of which employ the Adam optimizer and multilayer perceptron networks to identify the superior architecture. The subsequent phase implements Convolutional Block Attention Modules (CBAM) using optimized structural configurations to refine feature representations and enhance computational effectiveness. Performance assessment encompasses multiple evaluation indicators, namely classification accuracy, precision rates, recall values, F1-measure, and receiver operating characteristic curve area (ROC-AUC). Through the integration of attention-based techniques and the application of pre-trained model knowledge, the framework demonstrates exceptional results using the DDSM database. The results indicated that pretrained model VGG-19 yielded the best accuracy and precision scores with data splitting ratio (80% train, 10% validation, and 10% test) where, the accuracy for testing achieve 97.75%, and precision achieve 100%. After integrating CBAM, the result improved to 99.99% for accuracy and 100% precision outperforming state-of-the-art methods.