Low-dose CT reconstruction using dataset-free learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Image Reconstruction

Background:

• Low-Dose Computed Tomography (LDCT) is crucial for reducing radiation exposure in medical imaging.
• Supervised deep learning methods for LDCT reconstruction require extensive paired low-dose and norm-dose training data, limiting practical application.
• Existing reconstruction algorithms often struggle with noise suppression and detail preservation in LDCT.

Purpose of the Study:

• To develop an unsupervised, training data-free method for enhancing LDCT image reconstruction.
• To eliminate the need for large datasets of paired low-dose and norm-dose CT images.
• To improve the quality of LDCT reconstructions by reducing noise and preserving fine structures.

Main Methods:

• A post-processing technique utilizing neural network training.
• Minimization of the ℓ1-norm distance between CT measurements and simulated sinogram data.
• Simultaneous minimization of the total variation (TV) of the reconstructed image, without requiring manual weight tuning.

Main Results:

• Effective suppression of noise in LDCT images.
• Preservation of fine anatomical structures.
• Demonstrated rapid convergence and low computational cost on AAPM challenge and LoDoPab-CT datasets.

Conclusions:

• The proposed unsupervised method offers a viable solution for high-quality LDCT reconstruction without extensive training data.
• This approach enhances the practical applicability of LDCT in clinical settings by improving image quality and reducing radiation risks.
• The method's efficiency and effectiveness in noise reduction and detail preservation are validated by experimental results.