Padé-ResNet: Improving the Accuracy and Stability of Medical Image Classification

Deep learning models, particularly convolutional neural networks (CNNs), have achieved remarkable results in the field of medical imaging. However, this domain demands exceptionally high precision, reliability, and interpretability, posing significant challenges for traditional CNNs. These challenges are partly due to the limitations of activation functions like ReLU, whose discontinuous nature impacts network robustness and stability, and increases the risk of "dying neurons". To address these issues, This study proposes a novel activation function based on Padé approximation, theoretically proving its capability to approximate any activation function, and also presents the corresponding learning algorithm. Building on this innovation, we designed the Padé-ResNet architecture. Experiments on the MedMNIST dataset demonstrate that Padé-ResNet not only delivers superior overall performance but also exhibits greater resilience against the Fast Gradient Sign Method (FGSM).