Paired Augmentation for Improved Image Classification using Neural Network Models

The lack of sufficient data points and class imbalances in datasets is a serious problem that mitigates against the success of deep learning classification models. These problems result in model overfitting, poor accuracy, and poor generalization. Traditional augmentation techniques, and advanced augmentation, such as adversarial approaches, have individually been found to be effective. In this work, we present a method to determine the most effective augmentation techniques to combine into the machine learning pipeline. We propose using only two augmentations in the pipeline, an advanced technique applied in an offline manner followed by a simple technique applied in an online manner. This approach is validated by application to two medical image problems using datasets characterized by class imbalances and small sizes. The former is a binary classification problem using a brain tumor dataset and the latter is a multi-label classification problem using a white blood cell dataset. Exhaustive experimentation with approximately 170 different combinations of augmentations methods is reported. Experimental results indicate a 15% and 11% improvement in validation accuracy over using no augmentation and a 2% improvement over using a single augmentation. We conclude that the pairing of augmentation methods results in improvements in the classification tasks.