Deformation-Recovery diffusion model (DRDM): Instance deformation for image manipulation and synthesis.

Area of Science:

• Medical Imaging
• Generative Models
• Computer Vision

Background:

• Diffusion models show promise for synthetic medical image generation.
• Existing methods often lack interpretable correspondences and can produce anatomically implausible results.
• There is a need for generative models that ensure anatomical plausibility and structural integrity.

Purpose of the Study:

• To propose a novel diffusion-based generative model, the Deformation-Recovery Diffusion Model (DRDM).
• To address limitations of current methods by emphasizing morphological transformation via deformation fields.
• To generate diverse, anatomically plausible deformations for improved medical imaging applications.

Main Methods:

• DRDM utilizes a topology-preserving deformation field generation strategy.
• It involves random sampling and integration of multi-scale Deformation Velocity Fields (DVFs).
• The model is trained to recover unrealistic deformation components, restoring deformed images to a realistic distribution.

Main Results:

• DRDM generates diverse, large-scale deformations while preserving anatomical plausibility.
• Experiments on cardiac MRI and pulmonary CT demonstrate the model's capabilities.
• Performance gains were observed in 2D image segmentation and 3D image registration tasks.

Conclusions:

• DRDM offers a novel approach to generative modeling in medical imaging.
• The model enhances data augmentation and synthesis, improving downstream task performance.
• DRDM shows significant potential for advancing medical image manipulation and analysis.