Advanced Ensemble Learning and Feature Enhancement for Robust Breast Cancer Classification in Histopathological Images

Breast cancer is one of the leading causes of death in women, making it important to detect early and correctly for proper treatment. The research uses the Breast Cancer Histopathological Image Dataset (BreaKHis) that contains 162 slide images scanned at 40x magnification and 277,524 image patches of 50x50 pixels, which were divided into two categories, IDC-negative and IDC-positive with 198,738 and 78,786 patches respectively. In the preprocessing stage, CIRES (Comprehensive Image Refinement and Enhancement System) is applied to improve image quality through contrast adjustment, noise reduction, normalization, and sharpening. Data augmentation is performed using CARSHI (Comprehensive Augmentation and Refinement System for Histopathology Images), applying rotation, flipping, zooming, and brightness adjustment to increase dataset diversity. Feature extraction is carried out using ContextAware Feature Enhancement, where the contextual relevance of feature is used for the adjustment of weights to improve discriminability of the feature. The Hybrid Convolutional Attention System identifies the most informative features in the case of feature selection. Classification is carried out by Ensemble Convolutional Recurrent Deep Network, a novel model which integrates models like LSTM, RNN, CNN, DNN, GRU, and DBN, employing bagging and boosting techniques for reduction of bias and variance. Hyperparameter tuning optimizes model performance, and the final prediction is computed by aggregating individual model outputs. E-CRDN achieved an accuracy of 96.5%, demonstrating strong performance across all evaluation metrics. This approach showcases its robustness and potential for reliable breast cancer classification, with significant application in clinical settings for accurate and efficient breast cancer detection.