Evolving Data Augmentation Strategies

Abstract

Deep Learning Algorithms are widely implemented and have reached state-of-the-art results in several scientific investigations. In medical images domain and Computer-Assisted Detection (CAD) systems, Convolution Neural Networks (CNNs) are the preferred deep network architecture. Despite getting good results, there are still some obstacles to overcome, namely the problem of overfitting. Lately, the Data Augmentation (DA) has been integrating the training pipeline of Deep Neural Networks to mitigate those issues. The effectiveness of classical image transformations in increasing the performance of classification tasks in medical imaging domain has been reported. However, the search for a suitable augmentation strategy is performed manually. This approach, mainly made of trial-and-error tasks can be very time-consuming and complex. Thereupon, a novel data augmentation approach is proposed. The approach is an Evolutionary Machine Learning approach that is able to automatically define an optimised DA strategy for each medical image classification task. Thus, the approach combines two algorithms, an evolutionary algorithm and combined with a deep learning algorithm to find suitable DA strategies. The results obtained demonstrate that the same or better level of performance is achieved when the Transformation Functions (TFs) and their parameters are defined automatically instead of manually.