Enhancing breast cancer detection and classification using machine learning and deep learning techniques

Abstract

Breast cancer remains one of the most prevalent and life-threatening diseases worldwide, newlinehighlighting the urgent need for improved diagnostic methodologies that offer newlinehigher accuracy, efficiency, and timeliness, particularly in early-stage detection. This newlineresearch investigates the potential of advanced deep learning architectures, enhanced newlinedata preprocessing, and augmentation strategies to address limitations such as class newlineimbalance, overfitting, and diagnostic delays. Leveraging the BreakHis dataset, comprising newline9,109 microscopic breast tumor images, a comparative analysis was conducted newlineon multiple convolutional neural network (CNN) models, including InceptionV3 V1, newlineDenseNet201, ResNet50, VGG16, and a hybrid CNN model. Data augmentation played newlinea pivotal role in improving generalization, with VGG16 and DenseNet201 achieving newlinetop accuracies of 91% and 90%, respectively, while the hybrid model, despite attaining newline99.83% training accuracy, showed signs of overfitting with a validation accuracy of newline69.53%. To further enhance diagnostic performance, a novel framework integrating a newlinenon-local means filter and Generative Adversarial Networks for image preprocessing newlinewas developed. The classification stage employed a Long Short-Term Memory network newlinewith Bidirectional Gated Recurrent Unit (BiGRU)-based Recurrent ShuffleNet newlineV2, combined with a Capsule Network and Graph Convolutional Neural Networks newline(CNGCNN), resulting in notable improvements in breast cancer stage identification newlinecompared to baseline methods. These gains were even more pronounced in disease progression newlinedetection, with increases of up to 4.5% in recall and 3.9% in accuracy, alongside newlinea 1.5% reduction in latency, enabling timely and targeted therapeutic interventions. newlineIn addition, this study explores an optimized multi-criteria ensemble learning approach newlineto further reduce processing time while maintaining CNN-level accuracy. DICOM images newlineunderwent preprocessing with SMOTE and data augmentation, followed by feature newlineextraction via CNN. The optimized images were then c