FDDM: Unsupervised Medical Image Translation with a Frequency-Decoupled Diffusion Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Diffusion models show promise for medical image translation but struggle with anatomical accuracy, especially with unpaired data.
• Accurate translation is crucial for disease diagnosis, localization, and treatment planning.

Purpose of the Study:

• To introduce the Frequency Decoupled Diffusion Model (FDDM) for high-fidelity MR-to-CT image conversion.
• To enhance the preservation of anatomical structures during medical image translation.

Main Methods:

• FDDM employs a two-stage approach: initial anatomical conversion followed by a guided diffusion process.
• A dual-path reverse diffusion process separates low-frequency and high-frequency information for improved image quality and anatomical fidelity.

Main Results:

• FDDM significantly outperformed existing GAN, VAE, and diffusion-based models on brain and pelvis MR-to-CT datasets.
• Achieved superior scores across FID, MAE, MSE, SSIM, and DICE metrics, notably excelling in FID.
• Demonstrated high-quality image generation while maintaining anatomical structure accuracy.

Conclusions:

• FDDM offers a robust solution for MR-to-CT conversion, balancing image quality with anatomical precision.
• The model facilitates more accurate downstream applications such as anatomy segmentation and radiotherapy planning.