Learning-based multi-material CBCT image reconstruction with ultra-slow kV switching.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Materials Science

Background:

• Spectral computed tomography (CT) imaging offers enhanced material differentiation compared to conventional CT.
• Multi-material decomposition is crucial for quantitative analysis in medical imaging.
• Ultra-slow kilovoltage (kV) switching techniques present unique challenges for spectral CT data acquisition.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for multi-material decomposition in spectral cone-beam CT (CBCT) imaging.
• To enable accurate material density reconstruction from ultra-sparse spectral CBCT projections acquired with ultra-slow kV switching.
• To assess the performance of the proposed method in distinguishing various materials, including iodine and bone.

Main Methods:

• A novel deep neural network, SkV-Net, was developed, featuring a U-Net backbone and a multi-head axial attention module.
• SkV-Net processes CT images reconstructed from each kV acquired during ultra-slow kV switching.
• The network outputs basis material images by leveraging energy-dependent attenuation characteristics.

Main Results:

• The SkV-Net successfully reconstructed four material density images (fat, muscle, bone, iodine) from sparse spectral projections.
• Physical experiments demonstrated decomposition errors of less than 6% for iodine and CaCl2.
• The results indicate high precision in material differentiation using the proposed deep learning approach.

Conclusions:

• SkV-Net presents a promising solution for multi-material decomposition in spectral CBCT imaging systems utilizing ultra-slow kV switching.
• The deep learning approach offers high accuracy and precision in material density reconstruction.
• This method has the potential to advance quantitative imaging in medical applications.