A Multi-stage Optimization Architecture for Effective Breast Cancer Diagnosis Based on Deep Neural Networks

Breast cancer ranks second in fatality among women. Conventional diagnostic methods are time-consuming, exhausting, and expensive, potentially leading to delayed treatment or misdiagnosis. Machine learning (ML) and deep learning (DL) methods have shown outstanding potential. However, they face challenges like feature identification and selection, data imbalance, and parameter optimization. Unlike most known solutions that have explored these stages for breast cancer prediction, either separately or in limited combinations, our approach simultaneously tackles these critical issues using multi-stage optimization architecture. Three well-established techniques, namely correlation analysis-based feature selection (CFS), LASSO regression, and mutual information (MI), are used for FS. Data balancing is performed using both oversampling and undersampling techniques, including the synthetic minority oversampling technique (SMOTE), k-nearest neighbor oversampling (KNNOR), and random undersampling (RUS). Finally, hyperparameter optimization (HPO) is carried out by adopting various methods including grid search, random search, Bayesian optimization, and semi-automatic to maximize the classification performance of seven renowned ML algorithms (logistic regression, decision tree, random forest, support vector machine, Naïve Bayes, k-nearest neighbor, and eXtreme gradient boosting), and a deep neural network (DNN). Through the experiments carried out on four publicly available datasets, including Wisconsin diagnostic breast cancer (WDBC), Wisconsin breast cancer dataset (WBCD), Wisconsin prognostic breast cancer (WPBC), and Breast cancer coimbra (BCC), the obtained results clearly demonstrate the superiority of the proposed method over the state-of-the-art methods.