Multi-planar dual adversarial network based on dynamic 3D features for MRI-CT head and neck image synthesis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy Physics

Background:

• Accurate electron density information is crucial for radiotherapy dose calculations.
• Magnetic Resonance Imaging (MRI) lacks electron density data, posing a challenge for dose calculation in radiotherapy planning.
• Current methods for generating CT data from MRI are insufficient for precise radiotherapy planning.

Purpose of the Study:

• To develop a generative adversarial network (GAN) capable of synthesizing Computed Tomography (CT) images from T1-weighted MRI acquisitions.
• To improve multimodal image synthesis quality for head and neck cancer patients.
• To enhance the accuracy of electron density information for radiotherapy dose calculations.

Main Methods:

• Proposed a Dual branch generator based on the U-Net architecture with an augmented multi-planar branch.
• The augmented branch extracts specific 3D dynamic features from volumetric MRI data.
• Employed an end-to-end convolutional U-Net embedding network for image synthesis.

Main Results:

• Achieved a Mean Absolute Error (MAE) of 18.76±5.167 HU in the sagittal plane and 26.83±8.27 HU for primary tumor regions in axial acquisitions.
• Obtained a Mean Structural Similarity (MSSIM) of 0.95±0.09 and a Frechet Inception Distance (FID) of 145.60±8.38 for sagittal synthesis.
• Demonstrated a 3.8% improvement over state-of-the-art GANs on a tumor test set, with superior peak signal-to-noise ratios (PSNR) for synthesized images.

Conclusions:

• The proposed dual CT synthesis model effectively produces high-quality synthetic CT (sCT) images from MRI data.
• The model's performance across various metrics indicates its superiority over existing state-of-the-art approaches.
• This technology has the potential to enhance clinical tumor analysis and radiotherapy treatment planning for head and neck cancers, pending further clinical validation.