Hierarchical multi-stage complementary network for medical image denoising

Abstract

Abstract

In computer vision, low-dose X-ray and CT images are often characterized by varying levels of quantum and electronic noise, which can adversely affect diagnostic accuracy. However, existing denoising algorithms may not fully address the specific noise characteristics of low-dose X-ray and CT images. To address this issue, we propose a Hierarchical Multi-stage Complementary Network (HMCNet) for low-dose X-ray and CT image denoising, which implements a complementary learning process that progressively aggregates local features, models global semantics, and refines local details. First, a hybrid noise module is constructed to generate paired noisy-original images at varying dose levels, which helps alleviate the scarcity of accurately paired medical image data. Second, a hierarchical feature learning network is designed to employ a multi-stage complementary learning mechanism for reducing signal intensity-dependent quantum and electronic noise. Finally, the proposed method is quantitatively assessed on X-ray and CT image datasets with established fidelity metrics, including Peak Signal-to-Noise Ratio and Structural Similarity Index. Experimental results on X-ray and CT image datasets demonstrate that the proposed method achieves effective noise suppression and notable image quality enhancement. Code is available at https://github.com/YDH130303/HMCNet .