Performance Analysis of Breast Cancer Classification using Different ML Classifiers and Activation Functions

Abstract

Accurate breast cancer detection using mammography is important for improving patient prognosis. However, manual interpretation remains challenged by tissue complexity and interobserver variability. Convolutional Neural Networks (CNNs) offer powerful automated feature extraction. However, their reliance on extensive annotated datasets limits their application in medical imaging. To address this, this study proposed a hybrid computer-aided diagnosis framework that combines high level deep features from CNN architecture with handcrafted Wavelet Scattering (WS) coefficients that are invariant to translation. This concatenated dual-feature representation is fed into multiple machine learning (ML) classifiers, specifically Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), and Ensemble learners, bypassing the traditional Softmax layer. Furthermore, the method is tested using ReLU and leakyReLU activation functions. The experimental results demonstrate that ensemble classifiers significantly outperform standalone models in handling high dimensional hybrid features. The proposed architecture utilizing an Ensemble Subspace k-NN classifier paired with a leakyReLU activation function achieved the highest testing accuracy of 91.60%. These findings establish that combining deep and handcrafted features with robust ensemble classification provides a highly resilient and accurate framework for mammographic mass classification.