Cross-Domain Unpaired Learning for Low-Dose CT Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised deep learning excels in low-dose computed tomography (LDCT) imaging but requires paired data, which is difficult to obtain clinically.
• The lack of paired data hinders the widespread clinical application of supervised deep learning for LDCT image enhancement.
• Generating realistic pseudo-paired LDCT data from unpaired datasets is challenging due to complex noise properties.

Purpose of the Study:

• To develop an effective cross-domain unpaired learning framework for generating pseudo LDCT data.
• To enable supervised deep learning for LDCT image reconstruction using readily available unpaired data.
• To improve the performance of LDCT imaging in clinical practice.

Main Methods:

• Proposed a cross-domain unpaired learning framework named CrossDuL.
• Developed a dedicated pseudo LDCT sinogram generative module utilizing a data-dependent noise model.
• Constructed a pseudo-paired dataset in the image domain (not sinogram domain) to train an LDCT image restoration module.

Main Results:

• The CrossDuL framework successfully generated pseudo LDCT data.
• Training on the pseudo-paired image dataset led to effective LDCT image reconstruction.
• Clinical datasets demonstrated promising quantitative and qualitative improvements in LDCT imaging performance.

Conclusions:

• The CrossDuL framework offers a viable solution for overcoming the paired data limitation in supervised deep learning for LDCT.
• The proposed method facilitates the generation of pseudo-paired data and subsequent image reconstruction.
• CrossDuL shows significant potential for enhancing clinical LDCT imaging quality and adoption.