Lightweight Super-Resolution for Low-Dose CT Images via Multi-Scale Feature Extraction

Low-dose computed tomography (LDCT) is extensively applied in clinical diagnosis, yet it suffers from high noise and low spatial resolution. A variety of methods have been developed to address these challenges, persistent limitations remain: (1) balancing reconstruction quality with model lightweightness; (2) insufficient feature extraction due to fixed-resolution features. Therefore, we propose a lightweight LDCT image super-resolution network via multiscale feature extraction. The model employs a hierarchical architecture and introduces a Hybrid Feature Extraction Block (HFEB) to acquire local details and global contextual information, refining reconstruction while maintaining efficiency. The HFEB includes a Local Convolution Block (LCB) for local texture extraction, a Multi-dimensional Feature Enhancement Block (MFEB) combining spatial and channel attention, and a Feature Refinement Attention Block (FRAB) for fine-grained feature enhancement. Additionally, a Hierarchical Reciprocal Attention Mixer (H-RAMi) is introduced for efficient multi-scale feature fusion and upsampling. We further improve the Gated-Dconv Feedforward Network (GDFN) from Restormer to reduce computational complexity without sacrificing quality. Experimental comparisons with five advanced methods on two public datasets indicate the effectiveness of the proposed method.