Virtual computed-tomography system for deep-learning-based material decomposition.

Area of Science:

• Medical Imaging
• Computational Biology
• Radiology

Background:

• Material decomposition (MD) in computed tomography (CT) is crucial for correlating anatomical and functional imaging.
• Accurate elemental composition of human tissues is a significant challenge in current MD techniques.
• Existing methods struggle with the complexity of real human body composition.

Purpose of the Study:

• To develop a virtual CT system model for generating realistic CT datasets with accurate elemental information.
• To create a deep learning (DL) model for precise material distribution estimation.
• To evaluate the performance of the DL model across different CT energy levels and its robustness to artifacts.

Main Methods:

• Generation of virtual CT datasets using a novel generative CT model based on ICRP110 human phantoms.
• Training a deep learning model to perform material decomposition using these synthetic datasets.
• Analysis of the DL model's accuracy and robustness using quad-, dual-, and single-energy CT data, including variations in beam-hardening artifacts, noise, and spectrum.

Main Results:

• The deep learning approach achieved precise material decomposition, even with single-energy CT images.
• The method demonstrated minimal impact from noise, beam-hardening artifacts, and spectrum variations.
• Accurate elemental information was successfully estimated for human phantoms.

Conclusions:

• The virtual CT system overcomes the challenge of preparing large CT databases for MD.
• The proposed DL-based technique offers a robust solution for accurate material decomposition.
• This technique holds significant potential for applications in clinical radiodiagnosis and radiotherapy.