A Correlative Survey on Imbalanced Breast Cancer Data Using Ensembled Oversampling Techniques and Deep Learning Algorithms

Inadequacy of labeled breast cancer data causes imbalance and inaccuracy in classifiers. Our proposed research aims for a computer assisted detection (CAD) system to detect maligned cancer cells using the Wisconsin Breast Cancer dataset through a novel ensembled approach that combines various oversampling methods with standard machine learning and deep learning classifiers. The work is three-folds: (1) Use of 6 oversampling techniques: Random oversampling, Synthetic minority oversampling technique, Borderline SMOTE, k-means SMOTE, Support vector machine SMOTE, Adaptive synthetic minority oversampling, (2) Use of 5 ML models: Logistic Regression, Decision Tree, Random Forest, Adaptive boosting, Extreme gradient boosting, and (3) Use of 3 DL models: Artificial neural network, Multilayer perceptron, Convolutional neural network (CNN). The experimental results deduce that CNN has performed better than the others when assisting the SMOTE family and achieves the maximum accuracy of 98.25%.