Number of Epochs of Each Model and Hyperband’s Classification Performance

Abstract

Computer-aided diagnosis (CAD) systems based on deep learning methods, such as the convolutional neural network (CNN), enable early breast cancer detection, diagnosis, and treatment. However, many studies based on CNNs usually train models by manually selecting various parameters, which is time-consuming and difficult to find the best solution. In this paper, we conceptualized a new, improved method to resolve these limitations. More specifically, we proposed a customized Hyperband hyperparameter tuner with increased epochs for hyperparameter tuning of CNNs for breast cancer whole mount slide image patch classification. Experimental results indicated that our Hyperband with increased epochs has better performance than Bayesian optimization and the original Hyperband tuner in terms of accuracy on the dataset called "Breast Histopathology Images" when computing time is sufficient and can resolve the performance stability issue of the original Hyperband tuner.