MMCA: Multi-stage multi-order context aggregation framework for LDCT denoising

Low-dose computed tomography (LDCT) is widely used to reduce patient radiation exposure, but this reduction often comes at the cost of increased noise in the CT images. Although various deep learning-based methods have been developed for LDCT denoising, most struggle to balance local perception and global contextual capture, thus failing to highlight valuable expressions. This paper presents a multi-stage multi-order context aggregation learning framework designed for high-resolution feature map. The framework combines local perception with adaptive context aggregation to improve performance. Each stage employs the macro-architecture of a vision transformer and integrates edge-enhancement features. Initially, the input passes through feature embedding blocks, followed by the stacking of multiple multi-order context aggregation modules to enable efficient feature interaction. The context aggregation modules effectively generate more discriminative representations from features that incorporate edge information. Extensive experiments on two publicly available LDCT denoising datasets demonstrate that our method surpasses state-of-the-art models. Notably, our method strikes a better balance between network efficiency and denoising performance. The code will be made publicly available on https://code.ihep.ac.cn/lijf/MMCA.