Main Article Content

A statistically validated and interpretable machine learning framework for breast cancer classification using shape descriptors


K. A. Sotonwa
J. O. Akintayo
A. Pearson
A. Sofoluwe
S. Arekete
S. Sammet
F. Olopade
B. Aribisala

Abstract

Breast cancer (BC) remains a major global health priority, with nearly half a million invasive cases reported in 2024. Early diagnosis is critical, and Machine Learning (ML) offers significant potential for automated detection. This study evaluated ML models using tumour shape descriptors from the Wisconsin Diagnostic Breast Cancer (WDBC) dataset (569 cases: 357 benign, 212 malignant). Ten numerical descriptors, including radius, texture, and concave points, were extracted from digitized fine needle aspirates (FNA). The dataset was split 70:30 into training and test sets, with Chi-square feature selection, mutual information (MI), and ANOVA feature selection applied to reduce dimensionality and prevent data leakage. Three models: Artificial Neural Network (ANN), Classification and Regression Tree (CART), and Logistic Regression (LR) were trained using stratified cross-validation and evaluated on the independent test set. LR achieved the highest ROC (97.31%), significantly outperforming ANN (p = 0.009) and CART (p = 0.001) by DeLong’s test, while McNemar’s test revealed no significant differences in misclassification rates (p > 0.05). CART achieved the best F1-score, reflecting balanced sensitivity and specificity. These findings demonstrated that statistically validated and interpretable ML models provide robust diagnostic performance for BC classification.


Journal Identifiers


eISSN: 2437-2110
print ISSN: 0189-9546