Enhancing Classification Performance in Knee Magnetic Resonance Imaging Using Adversarial Data Augmentation

The utilization of adversarial data augmentation has demonstrated the potential capability to enhance the classification performance in training deep neural networks to perform computer vision tasks. In this paper, we investigate the effectiveness of this approach as a strategy for enlarging a knee Magnetic Resonance Imaging (MRI) dataset by adding adversarial perturbation. Specifically, we use the Fast Gradient Sign Method (FGSM) to perturb a subset of the training dataset as extra training images to re-train a baseline model that was trained under the same configuration as the top-ranked model on the MRNet leaderboard. Particularly, unlike most of the current work, we investigate the impact of two hyperparameters (attack magnitude and the proportion of data to be re-trained) on the performance of area under the ROC curve (AUC), accuracy, sensitivity, and specificity. Additionally, our results show that adversarial data augmentation can further improve the well-trained baseline model's AUC by 0.26%, as well as provide a slight improvement in specificity at the same classification threshold. These findings underscore the potential advantages of adversarial data augmentation as a technique for optimizing the decision boundaries of deep learning models. The code of this work will be available on GitHub after the paper is published.