TransCT: Dual-Path Transformer for Low Dose Computed Tomography

Low dose computed tomography (LDCT) has attracted more and attention in routine clinical diagnosis assessment, therapy planning, etc., which can reduce the of X-ray radiation to patients. However, noise caused by low exposure degrades CT image quality then affects accuracy. In this paper, we train a transformer-based neural network enhance final quality. To be specific, first decompose noisy LDCT into two parts: high-frequency (HF) low-frequency (LF) compositions. Then, extract content features (\(X_{L_c}\)) latent texture (\(X_{L_t}\)) from LF part, as well HF embeddings (\(X_{H_f}\)) part. Further, feed \(X_{L_t}\) \(X_{H_f}\) modified transformer with three encoders decoders obtain well-refined features. After that, combine these pre-extracted \(X_{L_c}\) encourage restoration high-quality images assistance piecewise reconstruction. Extensive experiments on Mayo dataset show that our method produces superior results outperforms other methods.