Adversarial training-based robust diagnosis method for lumbar disc herniation

Currently, lumbar spine diseases are becoming increasingly young, and with the aging of the population, clinical doctors are facing increasing pressure in detecting lumbar spine diseases. Therefore, an AI-based diagnosis system for lumbar spine diseases using nuclear magnetic images (MRI) has become a sustainable solution for early diagnosis. However, a large amount of work has shown the fragility of neural networks in unseen data distributions. Therefore, this paper proposes an adversarial training-based robust diagnosis method for lumbar disc herniation to address the fragility issue of deep models under specific small perturbations. By enhancing the robustness of the model to specific perturbations through adversarial training, the deep network can correctly classify lumbar spine MRI data with perturbations. The deep network model uses ResNet50, with adversarial examples containing adversarial perturbations added during training, followed by joint training of normal and adversarial examples, and Mixup augmentation from the perspective of data augmentation to further enhance the model's robustness. Through 5-fold cross-validation training, this method was verified to significantly improve the robustness of the model under adversarial perturbations (average recognition accuracy increased from 50.14% to 71.07%), while maintaining high recognition accuracy for normal samples (our method/baseline: 89.14%/89.05%).