MR-based synthetic CT generation using dual-attention enhanced 3D Conditional GAN for head and neck radiotherapy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy Planning

Background:

• Magnetic Resonance Imaging (MRI) offers superior soft-tissue contrast compared to Computed Tomography (CT).
• CT is essential for radiotherapy planning due to its accurate electron density information.
• Synthesizing CT from MRI (sCT) can enable MR-only radiotherapy planning, reducing radiation exposure and improving workflow.

Purpose of the Study:

• To develop and validate an improved 3D conditional generative adversarial network (3D cGAN) for synthesizing CT images from MRI data.
• To enhance the 3D cGAN with a dual-attention mechanism for improved accuracy in head and neck tumor radiotherapy planning.
• To assess the clinical acceptability and quantitative performance of the synthesized CT (sCT) images.

Main Methods:

• Utilized 212 paired CT and T1-weighted MRI datasets (180 public, 32 clinical).
• Implemented a 3D cGAN framework with structural modifications to generator, discriminator, and loss functions.
• Introduced a lightweight dual-attention mechanism module based on a 3D residual network into the generator.

Main Results:

• The dual-attention enhanced 3D cGAN successfully generated clinically acceptable sCT images for all 26 test cases.
• Achieved high quantitative accuracy: NCC 97.06%, SSIM 90.24%, PSNR 28.23 ± 0.42, MAE 32.53 ± 2.49 HU.
• Outperformed U-Net, CycleGAN, and basic 3D cGAN across all evaluated quantitative metrics.

Conclusions:

• The proposed dual-attention enhanced 3D cGAN algorithm effectively synthesizes CT images from MRI for head and neck tumor patients.
• This automated method is crucial for advancing MR-only radiotherapy planning.
• The approach offers rapid and accurate generation of synthetic CT, improving radiotherapy planning efficiency and safety.