Innovative Methods for Breast Cancer Diagnosis

Abstract

Breast cancer ranks as the second leading cause of cancer mortality in women globally and an early detection is crucial to enhance patient survival. In this work, we present a novel machine learning based diagnostic system that combines kernel-based dimension reduction and a number of individual and ensemble classifiers. Our approach exploits Kernel Principal Component Analysis (Kernel PCA) for nonlinear dimensionality decomposition, followed by classification by employing Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), Random Forest (RF), and XGBoost (XGB). To boost our performance more, we use 2 ensemble methods: Soft Voting based Ensemble to average the probabilistic result along all base-models and Improved Stacking based Ensemble which is an ensemble in which we fit the meta-model using Logistic Regression and passing the original features. Comprehensive experiments are conducted on the UCI Breast Cancer Wisconsin (Diagnostic) dataset, which show that ensembled models outperform standalone models with the classification accuracy of up to 96%. Testing metrics including ROC AUC, cross validation accuracy, and confusion matrices validate both the sensitivity and fitness of the proposed method. This work is part of the continuous effort to develop AI systems that are both interpretable, and that provide high performance AI in medical diagnostics.