Denoising Diffusion Wavelet Models for Zero-shot Medical Image Translation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Diffusion models show promise for image generation but struggle with medical image translation, particularly CBCT to CT, due to anatomical detail loss.
• Conventional methods like GANs and VAEs face challenges in bidirectional distribution mapping and robustness to out-of-distribution data.
• Accurate CBCT to CT translation is crucial for radiation therapy planning and diagnosis.

Purpose of the Study:

• To develop a novel diffusion model for high-fidelity CBCT to CT image translation.
• To enhance anatomical structure preservation during the translation process.
• To achieve robust performance on both in-distribution and out-of-distribution datasets.

Main Methods:

• Proposed a denoising diffusion wavelet model (DDWM) that learns only the CT data distribution.
• Implemented a similarity-bridge-controlled reverse diffusion process to fuse domain-invariant information from CBCT.
• Utilized wavelet transform to decompose images and identify similar frequency bands for preserving anatomical structures.

Main Results:

• DDWM demonstrated superior performance across multiple metrics (FID, PSNR, MAE, DICE scores) compared to state-of-the-art methods.
• The model achieved high-quality translation and preserved intricate anatomical details from the source CBCT images.
• Excellent results were observed on both in-distribution and out-of-distribution CBCT-to-CT translation datasets.

Conclusions:

• The proposed DDWM effectively addresses the limitations of existing methods for CBCT to CT translation.
• DDWM offers a robust and structure-faithful approach for medical image translation, outperforming conventional techniques.
• This method holds significant potential for improving diagnostic accuracy and treatment planning in medical imaging.