Developing a predictive model for breast cancer detection using radiomics-based mammography and machine learning

Abstract Background This study focuses on classifying benign and malignant breast lesions in mammography images using BI-RADS, combining ten feature selection and classification methods to improve diagnostic accuracy. These advanced techniques aim to overcome current challenges in breast cancer diagnosis. Results We evaluated 100 combinations of machine learning procedures using tenfold cross-validation to assess the stability of the radiomics model. The Extra Trees, Random Forest, and Gradient Boosting Decision Tree classification algorithms demonstrated the best diagnostic performance for mammography. The Principal Component Analysis (PCA) feature selection combined with the Extra Trees classifier (accuracy: 0.960 and AUC = 0.997), PCA with Random Forest (accuracy: 0.953 and AUC = 0.993), and PCA with Gradient Boosting (accuracy: 0.938 and AUC = 0.988) achieved the highest prediction accuracy and area under the curve (AUC). The PCA with Gradient Boosting classifier (sensitivity: 0.963) and Pearson Correlation with Gaussian Naïve Bayes (sensitivity: 0.957) showed the highest sensitivity. For specificity, PCA with Extra Trees (specificity: 0.990) and PCA with Random Forest (specificity: 0.977) reached the highest levels, respectively. Conclusions Promising classification results from this study demonstrated the effectiveness of machine learning based on radiomics features in differentiating between normal, benign, and malignant tissues, contributing to improved breast cancer diagnosis and treatment.