Image Decomposition Algorithm for Dual-Energy Computed Tomography via Fully Convolutional Network.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Dual-energy computed tomography (DECT) offers enhanced substance identification through spectral information.
• Accurate material-specific imaging in DECT relies heavily on effective basis material decomposition methods.

Purpose of the Study:

• To develop and validate a data-driven algorithm for image-based material decomposition in DECT.
• To address the critical need for improved accuracy and robustness in DECT image analysis.

Main Methods:

• A deep neural network architecture, incorporating a fully convolutional net (FCN) and a fully connected net, was proposed.
• The FCN extracts image features, while the fully connected net computes decomposed basic material coefficients.
• The model was trained and validated using a modified clinical dataset.

Main Results:

• The proposed FCN demonstrated a 60% reduction in bias and a 70% decrease in standard deviation compared to existing algorithms.
• The algorithm exhibits superior material separation capabilities.
• Excellent performance was maintained even in the presence of photon noise.

Conclusions:

• The developed deep cascaded network achieves high decomposition accuracy and robust noise performance.
• The study highlights the strong function-fitting capabilities of deep neural networks for DECT problems.
• Deep learning presents a promising avenue for solving complex, nonlinear challenges in DECT imaging.