Unpaired low-dose computed tomography image denoising using a progressive cyclical convolutional neural network.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) reduces patient radiation risk but suffers from noise that impedes diagnosis.
• Deep learning excels at LDCT denoising, yet most methods require paired normal-dose and low-dose CT images, which are clinically scarce.
• Unsupervised methods offer a more generalizable approach to LDCT denoising.

Purpose of the Study:

• To develop a simpler, memory-efficient unsupervised deep learning framework for LDCT denoising using unpaired data.
• Introduce the progressive cyclical convolutional neural network (PCCNN) for noise removal in the latent space of CT images.

Main Methods:

• Proposed a progressive cyclical convolutional neural network (PCCNN) trained on unpaired low-dose CT (LDCT) and normal-dose CT (NDCT) images.
• Incorporated a noise transfer model to transfer noise from LDCT to NDCT.
• Utilized a multi-stage wavelet transform within a progressive module to remove noise while preserving high-frequency details like edges.

Main Results:

• PCCNN demonstrated superior denoising performance compared to seven other LDCT algorithms on the AAPM dataset.
• Achieved significant improvements in peak signal-to-noise ratio (PSNR) from 29.622 to 30.671 and structural similarity index (SSIM) from 0.8544 to 0.9199.
• Qualitative results showed enhanced resolution, detail preservation, and reduced artifacts, with visual assessments indicating a balance in noise suppression, contrast retention, and lesion discrimination.

Conclusions:

• The proposed PCCNN scheme achieves reconstruction results comparable to supervised learning methods.
• Demonstrated excellent performance in image quality and medical diagnostic acceptability for unsupervised LDCT denoising.
• PCCNN offers a viable clinical solution for improving LDCT image analysis without requiring paired data.