A feature invariant generative adversarial network for head and neck MRI/CT image synthesis

Area of Science:

• Medical Imaging
• Radiotherapy Planning
• Artificial Intelligence in Healthcare

Background:

• Magnetic Resonance Imaging (MRI) workflows are increasingly vital in radiotherapy.
• Accurate dose planning necessitates Computed Tomography (CT) for bone attenuation calculations.
• Bridging the gap between MRI and CT data is crucial for advanced radiotherapy.

Purpose of the Study:

• To develop a novel unsupervised deep image synthesis model for generating CT images from MRI data.
• To enable robust radiotherapy dose planning using only MRI-derived synthetic CT (sCT) images.
• To improve the clinical workflow for MRI-based radiotherapy treatments.

Main Methods:

• A generative adversarial network (GAN) was employed for unsupervised CT image synthesis from MRI.
• A novel invariant representation was learned, encoding convolutional feature maps for robust image generation.
• Histogram matching and a multi-resolution framework were integrated to ensure synthetic CT (sCT) image quality and accuracy.

Main Results:

• The proposed model demonstrated high-quality sCT image synthesis, outperforming existing generative models.
• Quantitative evaluation showed a mean absolute error of 26.44 and a Hounsfield unit error of 45.3.
• A high overall Dice coefficient of 0.74 was achieved, indicating excellent agreement with ground-truth CT data.

Conclusions:

• The novel GAN-based model effectively generates synthetic CT images from MRI for radiotherapy planning.
• The approach enhances MR-based workflows by providing accurate dose attenuation information without traditional CT.
• This synthesis model shows significant potential for improving radiotherapy planning applications.