Efficient and Lightweight Block-Based Feature Selection Approach Using Dice Coefficient with Machine Learning for Accurate Breast Cancer Classification
International journal of intelligent engineering and systems · 2025
Breast cancer remains one of the most prevalent and life-threatening diseases affecting women worldwide.Machine learning (ML) techniques have increasingly been recognized for their potential in achieving early and precise diagnosis.Within this context, feature selection is essential for improving classification accuracy, minimizing overfitting, and enhancing the interpretability of the models.However, many existing filter-based feature selection methods, such as Principal Component Analysis (PCA), Pearson Correlation Coefficient (PCC), and statistical scoring techniques, either rely on assumptions of linearity or retain irrelevant features.This leads to increased computational cost and suboptimal accuracy.To address these limitations, this paper proposes a lightweight feature selection approach based on the Dice Coefficient, which evaluates the similarity between each feature and the class label.The method is simple, fast, and does not require iterative model training.It introduces a block-based mechanism, where samples are grouped into blocks ranging from 2 to 10 samples each, to better capture feature-label associations and enhance selection stability.The method was evaluated using four ML classifiers: SVM, DT, KNN, and RF.The proposed approach achieved a peak accuracy of 99.12% on the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, using only 12 selected features.It reduced memory usage to 14.6 MB and execution time to 1.22 seconds.Compared to using the full feature set, the method achieved up to 41.9% reduction in execution time and over 65% reduction in memory usage, without sacrificing accuracy.These results demonstrate that the Dice-based method significantly outperforms conventional filter, wrapper, and embedded methods in both accuracy and efficiency.Its high predictive performance and low computational demand make it a strong candidate for use in real-time and embedded medical diagnostic systems.