Texture-aware dual domain mapping model for low-dose CT reconstruction.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Deep learning has advanced low-dose computed tomography (CT) reconstruction.
• Preserving CT textures remains a challenge for low-dose CT (LDCT) reconstruction.
• Existing deep learning methods often focus on single domains, limiting texture perception.

Purpose of the Study:

• To develop a dual-domain (sinogram and image) method for enhanced texture perception in LDCT reconstruction.
• To drive the reconstruction process using visual effects for improved texture preservation.
• To overcome limitations of existing single-domain or basic dual-domain approaches.

Main Methods:

• A novel dilated residual network (S-DRN) was designed for the sinogram domain to capture multiscale information.
• A self-attention (SA) residual encoder & decoder network (SRED-Net) was developed for the image domain to enhance edges and textures.
• A composite loss function, including feature loss from a boundary and texture feature-aware network (BTFAN) and mean square error (MSE), was utilized.

Main Results:

• The proposed texture-aware dual domain mapping network (TADDM-Net) demonstrated state-of-the-art performance on both objective and visual metrics.
• Validation was performed using American Association of Physicists in Medicine (AAPM)-Mayo Clinic LDCT datasets and real clinical data.
• The method achieved superior denoising and texture restoration capabilities.

Conclusions:

• The TADDM-Net significantly improves the visual quality of reconstructed CT images compared to single-domain or existing dual-domain methods.
• The proposed approach offers superior texture preservation and artifact reduction.
• The study provides intuitive evidence for model interpretability in texture-aware reconstruction.