RDA-Net: Residual Dual Attention Network with ASPP for Multi-Scale Low-Dose CT Image Denoising

Abstract

Although low-dose CT (LDCT) helps to minimize radiation risks, it typically introduces noticeable noise artifacts that degrade the image clarity and hinder accurate clinical assessments. To address this problem, we introduce a new multi-level attention-based deep learning framework, RDA-Net, for efficiently suppressing noise in LDCT images and enhancing the recovery of structural details. The network, which has an encoder–decoder architecture, incorporates a composite convolutional module (Conv3×3→Conv5×5) to expand the receptive field. Temperature-enhance channel and special attention modules, which combine lightweight channel attention with temperature-regulated spatial attention to facilitate fine-grained feature modeling and noise suppression, are symmetrically embedded within the main network structure. This design effectively highlights critical regions while suppressing redundant features. To better recover fine structural details, the bottleneck layer incorporates an atrous spatial pyramid pooling module that captures contextual features through dilated convolutions at multiple scales. In addition, the network incorporates a three-level skip residual connection strategy to preserve shallow features, and also employs input-level residual learning to enhance training stability and reconstruction accuracy. Validation using the open-access LDCT dataset from the 2016 AAPM-Mayo Low Dose CT Challenge indicates that RDA-Net achieves values of 33.0235, 0.9100, and 9.1173 for the peak signal-to-noise ratio, structural similarity index measure, and root mean square error, respectively. RDA-Net demonstrates significant improvements over existing methods such as EDCNN, RED-CNN, and CTformer, and yields the best overall performance. Finally, experimental results confirm the model’s strong ability to suppress noise and preserve details, underscoring its applicability in real-world scenarios.