Laryngeal disease classification is a relatively hard task in medical image processing resulting from its complex structures and varying viewpoints in data collection. Some existing methods try to tackle this task via the convolutional neural network, but they more or less ignore the intrinsic difficulty differences among different input samples and suffer from high training complexity. In order to better resolve these problems, an end-to-end Hierarchical Dynamic Convolutional Network (HDCNet) is proposed, which can dynamically process the input samples based on their difficulty. For the easy-classified samples, the HDCNet processes them with a smaller resolution and a relatively small network, while the difficult samples are passed to a large network with a larger resolution for more accurate classification results. Furthermore, a Feature Reuse Module (FRM) is designed to transfer the features learned by the small network to the corresponding block in the deep network to enhance the overall performance of some rather complicated samples. To validate the effectiveness of the proposed HDCNet, comprehensive experiments are conducted on the public available laryngeal disease classification dataset and HDCNet provides superior performances compared with other current state-of-the-art methods.
© 2022. The Author(s).