Synthetic Low-Energy Monochromatic Image Generation in Single-Energy Computed Tomography System Using a...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Radiotherapy Physics

Background:

• Dual-energy CT (DECT) offers valuable energy-specific imaging, but its limited availability restricts its widespread clinical use.
• Single-energy CT (SECT) is the predominant imaging modality, necessitating methods to extract advanced information from it.
• Improving image quality in head and neck cancer imaging is crucial for accurate radiotherapy planning and patient outcomes.

Purpose of the Study:

• To develop and validate a novel transformer-based deep learning model (SwinUNETR) for generating synthetic low-energy virtual monochromatic images at 50 keV (sVMI50keV) from SECT data.
• To assess the performance of the SwinUNETR model in terms of image quality and accuracy compared to a conventional U-Net model.
• To demonstrate the potential of SECT-derived sVMI50keV for enhancing head and neck cancer imaging in facilities lacking DECT.

Main Methods:

• A transformer-based deep learning model, SwinUNETR, was trained using DECT data from 70 head and neck cancer patients.
• The trained model was used to generate sVMI50keV from SECT images of 15 additional patients with both SECT and DECT data.
• Image quality and accuracy were evaluated by comparing generated sVMI50keV with true VMI50keV derived from DECT, using mean absolute error and contrast analysis against a U-Net model.

Main Results:

• The SwinUNETR model achieved a lower mean absolute error (33.0 ± 4.4 HU) compared to the U-Net model (36.5 ± 4.9 HU) when generating sVMI50keV.
• SwinUNETR demonstrated superior accuracy in tissue attenuation values and generated contrast changes more closely resembling DECT-derived VMI50keV.
• The generated sVMI50keV from SECT using SwinUNETR showed improved image quality relevant for head and neck cancer imaging.

Conclusions:

• Transformer-based models like SwinUNETR show significant potential for generating high-quality synthetic low-energy VMIs from conventional SECT images.
• This method offers a practical and feasible approach to improve head and neck imaging, extending benefits of advanced CT techniques to more patients and facilities.
• The study highlights a promising solution for overcoming DECT accessibility limitations in clinical practice.