Transformer Encoder-Decoder with depth-wise separable convolution for Low-Dose CT Denoising

Low-dose computed tomography ( LDCT) images frequently suffer from noise and artifacts due to diminished radiation doses, challenging the diagnostic integrity of the images. We introduce an innovative approach, the Transformer Encoder-Decoder with Depth-wise Separable Convolution (ECformer), which incorporates depth-wise separable convolution into the Transformer architecture. This synergistic fusion of convolutional neural network (CNN) and Transformer capabilities notably enhances spatial information capture while concurrently mitigating computational demands. Specifically, depth-wise separable convolution supplants traditional linear projection within the Transformer, facilitating efficient learning of local contextual dynamics. Further, the adoption of an encoder-decoder architecture bolsters the hierarchical feature representation, with shortcut connections ensuring the preservation of detail. Empirical evaluations on the AAPM-Mayo Clinic Low-Dose CT Grand Challenge Dataset confirm that ECformer outperforms existing denoising methods.