Data augmentation for building an ensemble of convolutional neural networks


Bioimage classification is important in subcellular localization, accurate cell phenotype recognition, and histopathological classification, to name a few applications. In this paper, we propose an ensemble of deep learning methods built using different batch sizes, different learning rates, and different methods of data augmentation. Ourmain goal is to test different methods of data augmentation for building an ensemble that boosts the performance of Convolutional Neural Networks (CNN).Our method is evaluated on a diverse set of bioimage classification problems, with each represented by a benchmark dataset and with each bioimage classification task representing a typical cellular or tissue-level classification problem. The results on these datasets demonstrate that the proposed ensemble does indeed boost the performance of the standard CNN. The MATLAB code of all the descriptors and experiments reported in this paper is available at

Keywords Microscopy imaging classification; Deep learning; Convolutional neural networks; Support vector machines.

[full paper]