Abstract
Computed tomography (CT) represents a cornerstone of modern clinical diagnostics, yet the ionizing radiation inherent to standard scanning protocols poses non-trivial health risks to patients. Sparse-view CT acquisition significantly reduces the radiation burden by subsampling projection angles; however, conventional analytical reconstruction techniques like Filtered Back Projection (FBP) yield pronounced streak artifacts and diagnostic degradation under such severely underdetermined conditions. In this study, we propose an optimized deep learning framework that couples an encoder-decoder residual topology with a hybrid attention mechanism for high-fidelity sparse-view CT reconstruction. The novel Hybrid Spatial-Channel Attention (HSCA) block dynamically models long-range spatial dependencies along projection trajectories while adaptively recalibrating feature channel responses, effectively decoupling anatomical tissue structures from non-local streaking patterns. We evaluated our model on clinical thoracic and abdominal CT datasets across challenging sampling regimes (30, 60, and 90 projection views). Quantitative evaluations demonstrate that our framework achieves superior performance, exceeding baseline convolutional architectures by up to 3.42 dB in peak signal-to-noise ratio (PSNR) and attaining a structural similarity index measure (SSIM) of 0.968 under 60-view conditions. Furthermore, edge-preservation indices and visual qualitative assessments confirm that the proposed hybrid architecture preserves subtle pathological margins and micro-calcifications without introducing secondary hallucination artifacts, providing a computationally efficient solution directly deployable on modern medical imaging instrumentation.
