Low-Dose CT Image Reconstruction Using Uncertainty-Aware Convolution–Attention Networks

Abstract

Reducing radiation dose in computed tomography often leads to increased noise and loss of structural detail. Motivated by hybrid convolution–attention denoising models such as CTLformer, this work investigates an uncertainty-aware reconstruction framework that combines local convolutional filtering with global self-attention. The model incorporates uncertainty estimation to adaptively balance fine-grained texture preservation and noise suppression. Experiments are conducted on two public low-dose CT datasets containing over 45,000 paired normal- and low-dose slices. Comparisons are performed against CNN-based methods (RED-CNN, DnCNN), transformer-based models, and recent hybrid architectures. Quantitative results show average improvements of 0.9–1.3 dB in PSNR and 1.5%–2.4% in SSIM under standard low-dose settings, with reduced variance across anatomical regions.