MRCON-Net: Multiscale reweighted convolutional coding neural network for low-dose CT imaging.

Area of Science:

• Medical Imaging
• Radiology
• Deep Learning

Background:

• Low-dose computed tomography (LDCT) is crucial for minimizing radiation exposure.
• Reducing radiation dose in LDCT can degrade image quality, introducing noise and artifacts.
• Existing deep learning (DL) methods for LDCT image enhancement often lack interpretability and can lose image details.

Purpose of the Study:

• To develop an interpretable and effective DL framework for enhancing LDCT images.
• To address the challenges of noise reduction, artifact removal, and detail preservation in LDCT.
• To improve the diagnostic accuracy of LDCT by enhancing image quality.

Main Methods:

• A multiscale reweighted convolutional coding neural network (MRCON-Net) was developed.
• The network incorporates a reweighted convolutional sparse coding approach inspired by ISTA.
• Dilated convolutions extract multiscale features, and a channel attention mechanism refines feature weights.

Main Results:

• MRCON-Net effectively reduces artifact noise and preserves anatomical structures in LDCT images.
• Quantitative analysis on AAPM Challenge and UIH clinical datasets showed high SSIM (0.9491, 0.915) and PSNR (40.66, 42.44) values.
• The model demonstrated superior performance in enhancing LDCT image quality compared to existing methods.

Conclusions:

• The proposed MRCON-Net achieves significant improvements in subtle structure-enhanced LDCT imaging.
• Ablation studies confirmed the effectiveness of individual components in enhancing model performance.
• MRCON-Net offers a promising solution for high-quality LDCT imaging with reduced radiation dose.