- Research Article
- 10.29284/tb9nxq10
Robust Breast Cancer Classification Using Ensemble Feature Selection And Weighted Stability Regularization
- Jan 01, 2025
- INTERNATIONAL JOURNAL OF ADVANCES IN SIGNAL AND IMAGE SCIENCES
- Manmohan Sahoo + 1 more +1
Detecting breast cancer early can significantly improve patient survival and treatment outcomes. However, the large number of features in medical imaging data makes it challenging to identify the most relevant predictors. This study introduces a framework that strengthens breast-cancer diagnosis by combining several complementary feature-selection methods—Boruta, SHAP, Relief, a Multi-Objective Evolutionary Algorithm (MOEA), and an Autoencoder. The feature rankings from each method are combined using a weighted aggregation strategy, followed by a stability check to retain only consistently important features. The final set of stable features was tested on the Wisconsin Diagnostic Breast Cancer (WDBC) dataset by training several state-of-the-art classifiers—ExtraTrees, CatBoost, LightGBM, XGBoost, and a soft Voting Classifier—within a 10-fold cross-validation framework. All classifiers achieved over 94% accuracy and ROC–AUC above 0.98, with ExtraTrees performing best at approximately 95.4% accuracy and 0.991 ROC–AUC. Compared with training on all original features, the proposed pipeline raises accuracy by roughly 4.5 percentage points (from 91% to 95.44 %) and improves mean ROC–AUC from 0.96 to 0.991.SHAP analysis highlights worst concave points, worst area, worst radius, mean concavity and area error as the most influential features—findings consistent with established clinical evidence. By combining probabilistic, evolutionary and neural-network–based selection techniques, the framework embodies key principles of soft computing and provides a reliable, interpretable basis for computer-aided diagnosis. It can be readily extended to larger or more heterogeneous datasets and to other imaging-based disease-detection tasks.
Read more