BSRNet: Blind Super-Resolution of Low-Dose CT Images Based on Adaptive Routing

Conventional super-resolution (SR) methods for computed tomography (CT) typically rely on a fixed degradation model, such as downsampling and noise. However, the actual degradation process is more complex, making it difficult to accurately recover fine pathological details. To address this issue, we propose a blind super-resolution (BSR) method, named Blind Super-Resolution of Low-Dose CT Images Based on Adaptive Routing (BSRNet). BSRNet dynamically selects expert sub-networks in both spatial and channel dimensions according to the degradation level of the input image, effectively reducing noise and restoring details. We introduce a Depth-Separable Gated Feedforward Network (DGFN) that enhances the capture of both global and local features, improving sensitivity to lesion details. Additionally, we propose two fusion modules: the Dynamic Feature Fusion Module (DFFM) and the Adaptive Feature Fusion Module (AFFM). The DFFM replaces traditional U-Net skip connections to enable adaptive fusion of encoded and decoded features, while the AFFM dynamically adjusts the fusion ratio between shallow and deep features. Extensive experiments demonstrate that our approach outperforms existing SR methods across various datasets.