Protecting Lung CT Nodule Classification Models with Feature Fusion of Image and Pixel-level Features

Yunzheng Zhu, Aichi Chien, Yuan Tian

Abstract: Deep learning models have emerged as a powerful and cost-effective tool for medical image classification, particularly within oncology. Computer-aided nodule classification with lung CT is one of the most outstanding tasks because of its superior performance. However, it is still vulnerable to adversarial attacks, which could result in misdiagnosis in clinical practice. Adversarial training, one of the most popular and effective defense methods, has proven to be useful in defending against adversarial attacks and increasing the robustness of the model. However, adversarial training is quite time-consuming due to the requirement of additional training with adversarial attacked samples. Thus, we developed another online adversarial defense method, feature fusion, that does not require additional training but can still significantly improve the robustness of the model. Feature fusion is also compatible with any medical image classification model. The results show a significant reduction in the performance drop for three first-order gradient adversarial attacks, fast gradient sign method (FGSM), basic iterative method (BIM), and projected gradient descent (PGD). In addition, incorporating additional adversarial training after pretraining with the feature fusion-based classification method can further significantly strengthen the robustness of the model.

External IDs:dblp:conf/healthsec/ZhuCT24