Unsharp Structure Guided Filtering for Self-Supervised Low-Dose CT Imaging.

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Low-dose computed tomography (LDCT) imaging is crucial for reducing radiation exposure but faces challenges in image quality.
• Supervised deep learning methods for LDCT reconstruction require extensive high-quality reference data, limiting their clinical application.

Purpose of the Study:

• To develop a novel deep learning-based method for reconstructing high-quality LDCT images directly from low-dose projections without requiring clean reference images.
• To address the limitations of existing deep learning approaches in clinical LDCT imaging.

Main Methods:

• Proposed a novel Unsharp Structure Guided Filtering (USGF) method combining guided filtering and structure transfer using deep convolutional networks.
• Estimated structure priors from low-dose CT images using low-pass filters, which guided the reconstruction process.
• Incorporated traditional Filtered Back Projection (FBP) algorithms into self-supervised training for projection-to-image domain transformation.

Main Results:

• The USGF method demonstrated superior performance in noise suppression and edge preservation compared to existing methods across three datasets.
• Successfully reconstructed high-quality CT images directly from low-dose projections, mitigating the need for clean references.
• Structure priors effectively alleviated over-smoothing by transferring specific structural characteristics to the generated images.

Conclusions:

• The proposed USGF method offers a promising solution for high-quality LDCT image reconstruction without requiring clean references.
• USGF has the potential to significantly advance clinical applications of LDCT imaging by improving image quality and reducing artifacts.
• The integration of structure priors and self-supervised learning represents a novel approach for low-dose CT reconstruction.