On Hyperparameters Tuning for Deep Learning Segmentation Models Applied to COVID-19 Computerized Tomography Images

This paper presents a study of hyperparameters tuning for two different image deep learning semantic segmentation architectures applied to COVID-19 Computerized Tomography images, UNet and UNet++, using as encoder layer the feature extractors of five different Convolutional Neural Networks - ResNeXt, Xception, MobileNet, VGG13 with Batch Normalization and DenseNet121. The hyperparameters considered are the learning rate, number of epochs for training, batch size and the optimizer algorithm. In order to find the optimal set, instead of merely combining each possible set o hyperparameters which results in a high computational cost, it employs a Bayesian search algorithm with early stopping, leading to faster analysis and provides not only the hyperparameters optimal values but also the impact of each one of them over the resulting mean of Intersection over Union - metric chosen to evaluate the models performance - as well as the segmentation results. All images are available as a public anonymized datasets.