Improving Robustness of Medical Image Diagnosis with Denoising Convolutional Neural Networks

Convolutional neural networks (CNNs) are vulnerable to adversarial noises, which may result in potentially disastrous consequences in safety or security sensitive systems. This paper proposes a novel mechanism to improve the robustness of medical image classification systems by bringing denoising ability to CNN classifiers with a naturally embedded auto-encoder and high-level feature invariance to general noises. This novel denoising mechanism can be adapted to many model architectures, and therefore can be easily combined with existing models and denoising mechanisms to further improve robustness of CNN classifiers. This proposed method has been confirmed by comprehensive evaluations with two medical image classification tasks.