HybridFormer: a convolutional neural network-Transformer architecture for low dose computed tomography image denoising

Low-dose computed tomography (CT) is a potent strategy to minimize X-ray radiation and its detrimental effects on patients. However, reducing radiation significantly boosts noise in reconstructed images, causing blur and obscuring critical tissue details. This obscurity poses significant challenges for doctors in making accurate diagnoses. Traditional techniques like sinogram domain filtration and iterative reconstruction algorithms require inaccessible raw data. Thus, this article introduces HybridFormer, a revolutionary image-denoising model utilizing the Residual Convolution-Swin Transformer Network, designed to enhance images while preserving vital details. Firstly, this algorithm constructs residual convolution for local feature extraction and Swin Transformer for global feature extraction, boosting denoising efficacy. Secondly, to address texture detail errors, we introduced a combined attention transformer unit (CATU) with a cross-channel attentive fusion layer (CCAFL), integrated with residual blocks to form a residual convolution and Swin Transformer Fusion Block (RSTB). Finally, using RSTB, we developed a deep feature refinement module (DFRM) to preserve image details. To avoid smoothing, we combined multi-scale perceptual loss from ResNet-50 with Charbonnier loss into a composite loss function. Validated on the AAPM2016 Mayo dataset, HybridFormer outperformed other state-of-the-art algorithms, achieving improvements of 0.02 dB, 0.16%, and 0.28% in PSNR, SSIM, and FSIM, respectively. Compared with other advanced algorithms, the proposed algorithm achieved the best performance indicators, confirming its superiority.