Dual residual convolutional neural network (DRCNN) for low-dose CT imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) uses high radiation doses, posing health risks.
• Low-dose CT (LDCT) reduces radiation but introduces image artifacts and noise, hindering diagnosis.

Purpose of the Study:

• To develop an advanced deep learning model for effective low-dose CT image reconstruction.
• To enhance image quality and diagnostic accuracy in LDCT by suppressing noise and artifacts.

Main Methods:

• A dual residual convolution neural network (DRCNN) was designed for direct sinogram-to-image reconstruction.
• The DRCNN integrates analytical domain transformations and employs simultaneous feature extraction in sinogram and image domains.
• Residual shortcut networks were utilized for noise suppression and error reduction.

Main Results:

• The DRCNN effectively reduced sinogram noise while preserving crucial structural information.
• Experimental results showed superior performance compared to existing methods like RED-CNN and DP-ResNet.
• Peak Signal-to-Noise Ratio (PSNR) improvements of approximately 3 dB and 1 dB were achieved over RED-CNN and DP-ResNet, respectively.

Conclusions:

• The proposed DRCNN framework offers a significant advancement in LDCT image reconstruction.
• This method enhances image quality and diagnostic reliability for low-dose CT scans.
• DRCNN presents a promising solution for safer and more effective medical imaging.