Regularizing deep neural networks for medical image analysis with augmented batch normalization

Shengqian Zhu, Chengrong Yu, Junjie Hu

Abstract: Highlights•We present a straightforward and efficient technique called NABN, which can be a more powerful substitute for BN to enhance the generalization ability and robustness of neural networks.•NABN maintains the regularization effect of BN and enjoys the benefits of adding noise, e.g., smoothing the structure of the input space.•Rigorous experiments conducted on the medical image classification and segmentation tasks provide compelling evidence for the effectiveness of NABN. Visualization results further illustrate that the NABN helps the network locate complicated lesions.