Image-domain material decomposition for dual-energy CT using unsupervised learning with data-fidelity loss

Area of Science:

• Medical Imaging
• Computed Tomography
• Artificial Intelligence

Background:

• Dual-energy computed tomography (DECT) is crucial for quantitative medical imaging.
• Material decomposition in DECT is prone to noise amplification, degrading image quality.
• Existing methods struggle with noise suppression and require extensive training data.

Purpose of the Study:

• Develop an unsupervised learning framework for DECT material decomposition.
• Improve image signal-to-noise ratio (SNR) through data-measurement consistency.
• Address limitations of supervised learning in clinical settings.

Main Methods:

• Combined iterative decomposition with a deep learning-based image prior in a generative adversarial network (GAN).
• Incorporated a data-fidelity loss for measurement consistency in the generator.
• Trained a discriminator to distinguish low-noise from high-noise material-specific images.

Main Results:

• Reduced standard deviation (SD) in decomposed images by up to 97% in simulations and 95% in clinical data.
• Achieved structural similarity index measures (SSIMs) > 0.95 against ground truth in phantom studies.
• Demonstrated effective noise suppression and accurate decomposition without paired data.

Conclusions:

• Noise amplification in DECT material decomposition is a significant clinical challenge.
• The proposed unsupervised method achieves accurate material decomposition with efficient noise suppression.
• The framework eliminates the need for paired ground-truth data, facilitating clinical application.