A Comparative Performance of Genetic Algorithm and Bayesian Optimization for Hyperparameter Tuning for Mammogram Classification

Abstract

An accurate breast cancer classification utilizing Convolutional Neural Network (CNN) requires the best option of hyperparameter selection to create a robust and adaptive algorithm based on different datasets. Standard optimization algorithms are subjected to nondeterministic and restricted to integer-valued parameters that cause a restricted optimization process on a highly non-linear dataset such as mammogram images. In this study, hyperparameter tuning through two optimization methods, Genetic Algorithm optimization (GAO) and Bayesian optimization (BO), are compared based on the evaluation for breast mass classification of benign and malignant on a publicly available mammogram image of the INbreast dataset. The best model shows an increase in testing accuracy at 90.05% and balancing of sensitivity to the specificity of 0.803 to 0.9481, improving its true positive rate when optimized using the GAO method. The optimization process allows for the combination of genetic mutations of the parent and fusion improves the creation of a population for the best-trained network.