Multi-scale dilated dense reconstruction network for limited-angle computed tomography.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Limited-angle computed tomography (CT) reconstruction often produces images with severe artifacts due to incomplete projection data, blurring essential structural details.
• Existing methods struggle to effectively reconstruct high-quality images, particularly within narrow scanning angular ranges.

Purpose of the Study:

• To develop a novel deep learning approach to enhance limited-angle CT reconstruction performance.
• To specifically address artifact reduction and improve structural detail recovery for narrow scanning angles.

Main Methods:

• A deep learning-based iterative framework, the multi-scale dilated dense reconstruction network (MSDDRNet), was proposed.
• MSDDRNet employs a multi-scale dilated dense convolution neural network (MSDD-CNN) integrated with conventional algorithms.
• The network incorporates DenseNet-Like structures for feature enhancement and projection domain data constraints within an iterative process. Pre-training and model migration strategies were used to accelerate training.

Main Results:

• Numerical experiments showed MSDDRNet effectively corrects artifacts and reduces noise in limited-angle CT reconstructions.
• The method demonstrated superior structure recovery compared to existing techniques for limited scan angles.
• The framework was successfully extended to general scanning conditions and applications like dental CT.

Conclusions:

• MSDDRNet offers a robust solution for improving image quality in limited-angle CT.
• The proposed method shows promise for enhancing detail recovery and artifact reduction in challenging CT scenarios.
• This general framework has potential applications in other CT variations, including low-dose, sparse-data, and spectral CT.